[jira] [Commented] (LUCENE-8943) Incorrect IDF in MultiPhraseQuery and SpanOrQuery
[ https://issues.apache.org/jira/browse/LUCENE-8943?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16905178#comment-16905178 ] Christoph Goller commented on LUCENE-8943: -- I agree, we cannot realistically approximate the doc freq of phrases. And yes, actually the scoring problem I brought up is a kind of synonym issue. Usually, if we are using synonyms we want to score exact query matches higher than synonym matches. That's probably one of the reasons why SynonymQuery allows to specify boosts. I am having lots of multiword synonyms. W2k16 e.g. is a synonym for "Windows Server 2016". Different boosts for multiword synonyms don't work reliably since matches for "Windows Server 2016" may score much higher than those of W2k16 due to huge IDFs. I am not so much looking for an optimal BM25 scoring for Phrases / Multiphrases / Spans. Instead I am looking for a way to score them within BM25 so that boosts work as expected. One step into this direction would be to limit IDF values in case of Phrases / Multiphrases / Spans. In BM25 it seems to be very important that IDF saturates and currently the behavior of Phrases / Multiphrases / Spans contradicts that. With the solution I proposed we can get rid of huge IDF values for Phrases / Multiphrases / Spans. Therefore I still think we should do it. Plus it would make scores more camparable and boosts would work more reliable. Your post made me think of the problem in another way. If we had something like MultiWordsSynonymQuery, we could have even more control. Similar to SynonymQuery we could use one IDF value for all synonyms. Synonym boost would work much more reliably. MultiWordsSynonymQuery could be very general. In my last post I suggested to approximate docFreq instead of IDFs in order to gurantee saturation. For implementing it, I thought about adding a member variable pseudoStats (TermStatistics) to Weight, which would be used for computing SimScorer. Usually the values for pseudoStats would be computed bottom up (SpanWeight, PhraseWeight) from the subqueries. But we could implement a general MultiWordsSynonymQuery as subclass of BooleanQuery (only allowing disjunction) which would set (adapt) pseudoStats in all its subweights (docFreq as max docFreq of all synonyms just as SynonymQuery currently does). > Incorrect IDF in MultiPhraseQuery and SpanOrQuery > - > > Key: LUCENE-8943 > URL: https://issues.apache.org/jira/browse/LUCENE-8943 > Project: Lucene - Core > Issue Type: Bug > Components: core/query/scoring >Affects Versions: 8.0 >Reporter: Christoph Goller >Priority: Major > > I recently stumbled across a very old bug in the IDF computation for > MultiPhraseQuery and SpanOrQuery. > BM25Similarity and TFIDFSimilarity / ClassicSimilarity have a method for > combining IDF values from more than on term / TermStatistics. > I mean the method: > Explanation idfExplain(CollectionStatistics collectionStats, TermStatistics > termStats[]) > It simply adds up the IDFs from all termStats[]. > This method is used e.g. in PhraseQuery where it makes sense. If we assume > that for the phrase "New York" the occurrences of both words are independent, > we can multiply their probabilitis and since IDFs are logarithmic we add them > up. Seems to be a reasonable approximation. However, this method is also used > to add up the IDFs of all terms in a MultiPhraseQuery as can be seen in: > Similarity.SimScorer getStats(IndexSearcher searcher) > A MultiPhraseQuery is actually a PhraseQuery with alternatives at individual > positions. IDFs of alternative terms for one position should not be added up. > Instead we should use the minimum value as an approcimation because this > corresponds to the docFreq of the most frequent term and we know that this is > a lower bound for the docFreq for this position. > In SpanOrQuerry we have the same problem It uses buildSimWeight(...) from > SpanWeight and adds up all IDFs of all OR-clauses. > If my arguments are not convincing, look at SynonymQuery / SynonymWeight in > the constructor: > SynonymWeight(Query query, IndexSearcher searcher, ScoreMode scoreMode, float > boost) > A SynonymQuery is also a kind of OR-query and it uses the maximum of the > docFreq of all its alternative terms. I think this is how it should be. -- This message was sent by Atlassian JIRA (v7.6.14#76016) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Comment Edited] (LUCENE-8943) Incorrect IDF in MultiPhraseQuery and SpanOrQuery
[ https://issues.apache.org/jira/browse/LUCENE-8943?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16901076#comment-16901076 ] Christoph Goller edited comment on LUCENE-8943 at 8/6/19 1:54 PM: -- {{Thanks for your quick response Alan. I've been doing some thinking about adding up IDF values in case of simple phrase queries and I no longer think that is the way we should do it.}} {{The problem is that we can get very high IDF values, i.e. values that are considerably higher than the maximum IDF value for a single term!}} {{Consider an index with 10 million docs. The maximum IDF value (BM25) for a single term is 16.8. Assume we have 10 docs containing "wifi" and 10 docs containing "wi-fi" which is split by our tokenizer into 2 tokens. The IDF value for "wifi" will be 13.77. If we assume that "wi" and "fi" both occur only in "wi-fi" docs, we get an IDF of 27.5 for the "wi fi" phrase query which wee need in order to find our 10 "wi-fi" docs. If we search for wifi OR "wi fi" the docs containing "wi-fi" will score much higher!}} {{Admittedly, it is easy to construct examples in which adding the IDF values of phrase parts yields values that are too high. The assumption of independence of phrase parts does not normally apply. But BM25 has a saturation for IDF values and adding up IDF values breaks it. This seems to be a serious drawback.}} {{I propose to switch from combining IDF-values to calculating / approximating docFreq. For the OR-case SynonymQuery does this already. It uses the maximum. For the AND-case we could use something like}} {{docFreqPhrase = (docFreq1 * docFreq2) / docCount}} {{The intuition behind this is again independence of phrase parts. But by computing a docFreq we can guarantee the saturation for IDF.}} {{For the "wi fi" example we get docFreqPhrase of 10^-5 leading to an IDF of 16.8 (saturation) and the difference to the IDF of wifi is considerably smaller compared to adding up IDFs. If phrase parts are rare, we quickly run into saturation of the IDF. But we also get some reasonable values. Consider the phrase "New York". If we assume that 100,000 docs contain "new" and 10,000 docs contain "york". By applying the formula from above we get and IDF for the phrase "New York" of 11.5 which is roughly the number we get when we add up the IDFs of the parts (current Lucene behavior).}} {{We could even have some simple adjustments for the fact that usually the independence assumption is not correct. For both the OR-case and the AND-case we could make values a little bit higher. The exact way for approximating docFreq for the AND-case and the OR-case could be defined in the Similarity and it could be configurable.}} I also did some research with Google: {{(multiword OR N-gram) AND BM25 AND IDF}} Unfortunately I did not find anything that helps. {{Do you know about the benchmarks used to evaluate scoring in Lucene? Are there any phrase queries involved?}} {{Robert told me it’s very Trek-like, so probably no phrase queries?}} {{In my opinion something like BM25 can only get us to a certain level of relevance. Of course, we have to get it right. IDF values of phrases / SpanQueries should not have such a big effect on the score simply because they get too high IDF-values. We have to do something reasonable. But for real break-through improvements we need something like query segmentation or even query interpretation and proximity of query terms in documents should have a high impact on the score. That's why I think it is important to integrate PhraseQueries and SpanQueries properly into BM25.}} was (Author: gol...@detego-software.de): {{Thanks for your quick response Alan. I've been doing some thinking about adding up IDF values in case of simple phrase queries and I no longer think that is the way we should do it.}} {{The problem is that we can get very high IDF values, i.e. values that are considerably higher than the maximum IDF value for a single term!}} {{Consider an index with 10 million docs. The maximum IDF value (BM25) for a single term is 16.8. Assume we have 10 docs containing "wifi" and 10 docs containing "wi-fi" which is split by our tokenizer into 2 tokens. The IDF value for "wifi" will be 13.77. If we assume that "wi" and "fi" both occur only in "wi-fi" docs, we get an IDF of 27.5 for the "wi fi" phrase query which wee need in order to find our 10 "wi-fi" docs. If we search for wifi OR "wi fi" the docs containing "wi-fi" will score much higher!}} {{Admittedly, it is easy to construct examples in which adding the IDF values of phrase parts yields values that are too high. The assumption of independence of phrase parts does not normally apply. But BM25 has a saturation for IDF values and adding up IDF values breaks it. This seems to be a serious drawback.}} {{I propose to switch from combining IDF-values to cal
[jira] [Comment Edited] (LUCENE-8943) Incorrect IDF in MultiPhraseQuery and SpanOrQuery
[ https://issues.apache.org/jira/browse/LUCENE-8943?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16901076#comment-16901076 ] Christoph Goller edited comment on LUCENE-8943 at 8/6/19 1:52 PM: -- {{Thanks for your quick response Alan. I've been doing some thinking about adding up IDF values in case of simple phrase queries and I no longer think that is the way we should do it.}} {{The problem is that we can get very high IDF values, i.e. values that are considerably higher than the maximum IDF value for a single term!}} {{Consider an index with 10 million docs. The maximum IDF value (BM25) for a single term is 16.8. Assume we have 10 docs containing "wifi" and 10 docs containing "wi-fi" which is split by our tokenizer into 2 tokens. The IDF value for "wifi" will be 13.77. If we assume that "wi" and "fi" both occur only in "wi-fi" docs, we get an IDF of 27.5 for the "wi fi" phrase query which wee need in order to find our 10 "wi-fi" docs. If we search for wifi OR "wi fi" the docs containing "wi-fi" will score much higher!}} {{Admittedly, it is easy to construct examples in which adding the IDF values of phrase parts yields values that are too high. The assumption of independence of phrase parts does not normally apply. But BM25 has a saturation for IDF values and adding up IDF values breaks it. This seems to be a serious drawback.}} {{I propose to switch from combining IDF-values to calculating / approximating docFreq. For the OR-case SynonymQuery does this already. It uses the maximum. For the AND-case we could use something like}} {{docFreqPhrase = (docFreq1 * docFreq2) / docCount}} {{The intuition behind this is again independence of phrase parts. But by computing a docFreq we can guarantee the saturation for IDF.}} {{For the "wi fi" example we get docFreqPhrase of 10^-5 leading to an IDF of 16.8 (saturation) and the difference to the IDF of wifi is considerably smaller compared to adding up IDFs. If phrase parts are rare, we quickly run into saturation of the IDF. But we also get some reasonable values. Consider the phrase "New York". If we assume that 100,000 docs contain "new" and 10,000 docs contain "york". By applying the formula from above we get and IDF for the phrase "New York" of 11.5 which is roughly the number we get when we add up the IDFs of the parts (current Lucene behavior).}} {{We could even have some simple adjustments for the fact that usually the independence assumption is not correct. For both the OR-case and the AND-case we could make values a little bit higher. The exact way for approximating docFreq for the AND-case and the OR-case could be defined in the Similarity and it could be configurable.}} {{I also did some research with Google: }} {{(multiword OR N-gram) AND BM25 AND IDF}} Unfortunately, I did not find anything that helps. {{Do you know about the benchmarks used to evaluate scoring in Lucene? Are there any phrase queries involved?}} {{Robert told me it’s very Trek-like, so probably no phrase queries?}} {{In my opinion something like BM25 can only get us to a certain level of relevance. Of course, we have to get it right. IDF values of phrases / SpanQueries should not have such a big effect on the score simply because they get too high IDF-values. We have to do something reasonable. But for real break-through improvements we need something like query segmentation or even query interpretation and proximity of query terms in documents should have a high impact on the score. That's why I think it is important to integrate PhraseQueries and SpanQueries properly into BM25.}} was (Author: gol...@detego-software.de): {{Thanks for your quick response Alan. I've been doing some thinking about adding up IDF values in case of simple phrase queries and I no longer think that is the way we should do it.}} {{The problem is that we can get very high IDF values, i.e. values that are considerably higher than the maximum IDF value for a single term!}} {{Consider an index with 10 million docs. The maximum IDF value (BM25) for a single term is 16.8. Assume we have 10 docs containing "wifi" and 10 docs containing "wi-fi" which is split by our tokenizer into 2 tokens. The IDF value for "wifi" will be 13.77. If we assume that "wi" and "fi" both occur only in "wi-fi" docs, we get an IDF of 27.5 for the "wi fi" phrase query which wee need in order to find our 10 "wi-fi" docs. If we search for wifi OR "wi fi" the docs containing "wi-fi" will score much higher!}} {{Admittedly, it is easy to construct examples in which adding the IDF values of phrase parts yields values that are too high. The assumption of independence of phrase parts does not normally apply. But BM25 has a saturation for IDF values and adding up IDF values breaks it. This seems to be a serious drawback.}} {{I propose to switch from combining IDF-value
[jira] [Comment Edited] (LUCENE-8943) Incorrect IDF in MultiPhraseQuery and SpanOrQuery
[ https://issues.apache.org/jira/browse/LUCENE-8943?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16901076#comment-16901076 ] Christoph Goller edited comment on LUCENE-8943 at 8/6/19 1:52 PM: -- {{Thanks for your quick response Alan. I've been doing some thinking about adding up IDF values in case of simple phrase queries and I no longer think that is the way we should do it.}} {{The problem is that we can get very high IDF values, i.e. values that are considerably higher than the maximum IDF value for a single term!}} {{Consider an index with 10 million docs. The maximum IDF value (BM25) for a single term is 16.8. Assume we have 10 docs containing "wifi" and 10 docs containing "wi-fi" which is split by our tokenizer into 2 tokens. The IDF value for "wifi" will be 13.77. If we assume that "wi" and "fi" both occur only in "wi-fi" docs, we get an IDF of 27.5 for the "wi fi" phrase query which wee need in order to find our 10 "wi-fi" docs. If we search for wifi OR "wi fi" the docs containing "wi-fi" will score much higher!}} {{Admittedly, it is easy to construct examples in which adding the IDF values of phrase parts yields values that are too high. The assumption of independence of phrase parts does not normally apply. But BM25 has a saturation for IDF values and adding up IDF values breaks it. This seems to be a serious drawback.}} {{I propose to switch from combining IDF-values to calculating / approximating docFreq. For the OR-case SynonymQuery does this already. It uses the maximum. For the AND-case we could use something like}} {{docFreqPhrase = (docFreq1 * docFreq2) / docCount}} {{The intuition behind this is again independence of phrase parts. But by computing a docFreq we can guarantee the saturation for IDF.}} {{For the "wi fi" example we get docFreqPhrase of 10^-5 leading to an IDF of 16.8 (saturation) and the difference to the IDF of wifi is considerably smaller compared to adding up IDFs. If phrase parts are rare, we quickly run into saturation of the IDF. But we also get some reasonable values. Consider the phrase "New York". If we assume that 100,000 docs contain "new" and 10,000 docs contain "york". By applying the formula from above we get and IDF for the phrase "New York" of 11.5 which is roughly the number we get when we add up the IDFs of the parts (current Lucene behavior).}} {{We could even have some simple adjustments for the fact that usually the independence assumption is not correct. For both the OR-case and the AND-case we could make values a little bit higher. The exact way for approximating docFreq for the AND-case and the OR-case could be defined in the Similarity and it could be configurable.}} {{I also did some research with Google: (multiword OR N-gram) AND BM25 AND IDF}} Unfortunately, I did not find anything that helps. {{Do you know about the benchmarks used to evaluate scoring in Lucene? Are there any phrase queries involved?}} {{Robert told me it’s very Trek-like, so probably no phrase queries?}} {{In my opinion something like BM25 can only get us to a certain level of relevance. Of course, we have to get it right. IDF values of phrases / SpanQueries should not have such a big effect on the score simply because they get too high IDF-values. We have to do something reasonable. But for real break-through improvements we need something like query segmentation or even query interpretation and proximity of query terms in documents should have a high impact on the score. That's why I think it is important to integrate PhraseQueries and SpanQueries properly into BM25.}} was (Author: gol...@detego-software.de): {{Thanks for your quick response Alan. I've been doing some thinking about adding up IDF values in case of simple phrase queries and I no longer think that is the way we should do it.}} {{The problem is that we can get very high IDF values, i.e. values that are considerably higher than the maximum IDF value for a single term!}} {{Consider an index with 10 million docs. The maximum IDF value (BM25) for a single term is 16.8. Assume we have 10 docs containing "wifi" and 10 docs containing "wi-fi" which is split by our tokenizer into 2 tokens. The IDF value for "wifi" will be 13.77. If we assume that "wi" and "fi" both occur only in "wi-fi" docs, we get an IDF of 27.5 for the "wi fi" phrase query which wee need in order to find our 10 "wi-fi" docs. If we search for wifi OR "wi fi" the docs containing "wi-fi" will score much higher!}} {{Admittedly, it is easy to construct examples in which adding the IDF values of phrase parts yields values that are too high. The assumption of independence of phrase parts does not normally apply. But BM25 has a saturation for IDF values and adding up IDF values breaks it. This seems to be a serious drawback.}} {{I propose to switch from combining IDF-values to cal
[jira] [Commented] (LUCENE-8943) Incorrect IDF in MultiPhraseQuery and SpanOrQuery
[ https://issues.apache.org/jira/browse/LUCENE-8943?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16901076#comment-16901076 ] Christoph Goller commented on LUCENE-8943: -- {{Thanks for your quick response Alan. I've been doing some thinking about adding up IDF values in case of simple phrase queries and I no longer think that is the way we should do it.}} {{The problem is that we can get very high IDF values, i.e. values that are considerably higher than the maximum IDF value for a single term!}} {{Consider an index with 10 million docs. The maximum IDF value (BM25) for a single term is 16.8. Assume we have 10 docs containing "wifi" and 10 docs containing "wi-fi" which is split by our tokenizer into 2 tokens. The IDF value for "wifi" will be 13.77. If we assume that "wi" and "fi" both occur only in "wi-fi" docs, we get an IDF of 27.5 for the "wi fi" phrase query which wee need in order to find our 10 "wi-fi" docs. If we search for wifi OR "wi fi" the docs containing "wi-fi" will score much higher!}} {{Admittedly, it is easy to construct examples in which adding the IDF values of phrase parts yields values that are too high. The assumption of independence of phrase parts does not normally apply. But BM25 has a saturation for IDF values and adding up IDF values breaks it. This seems to be a serious drawback.}} {{I propose to switch from combining IDF-values to calculating / approximating docFreq. For the OR-case SynonymQuery does this already. It uses the maximum. For the AND-case we could use something like}} {{docFreqPhrase = (docFreq1 * docFreq2) / docCount}} {{The intuition behind this is again independence of phrase parts. But by computing a docFreq we can guarantee the saturation for IDF.}} {{For the "wi fi" example we get docFreqPhrase of 10^-5 leading to an IDF of 16.8 (saturation) and the difference to the IDF of wifi is considerably smaller compared to adding up IDFs. If phrase parts are rare, we quickly run into saturation of the IDF. But we also get some reasonable values. Consider the phrase "New York". If we assume that 100,000 docs contain "new" and 10,000 docs contain "york". By applying the formula from above we get and IDF for the phrase "New York" of 11.5 which is roughly the number we get when we add up the IDFs of the parts (current Lucene behavior).}} {{We could even have some simple adjustments for the fact that usually the independence assumption is not correct. For both the OR-case and the AND-case we could make values a little bit higher. The exact way for approximating docFreq for the AND-case and the OR-case could be defined in the Similarity and it could be configurable.}} {{I also did some research with Google: (multiword OR N-gram) AND BM25 AND IDF}} {{Unfortunately, I did not find anything that helps. }} {{Do you know about the benchmarks used to evaluate scoring in Lucene? Are there any phrase queries involved?}} {{Robert told me it’s very Trek-like, so probably no phrase queries?}} {{In my opinion something like BM25 can only get us to a certain level of relevance. Of course, we have to get it right. IDF values of phrases / SpanQueries should not have such a big effect on the score simply because they get too high IDF-values. We have to do something reasonable. But for real break-through improvements we need something like query segmentation or even query interpretation and proximity of query terms in documents should have a high impact on the score. That's why I think it is important to integrate PhraseQueries and SpanQueries properly into BM25.}} > Incorrect IDF in MultiPhraseQuery and SpanOrQuery > - > > Key: LUCENE-8943 > URL: https://issues.apache.org/jira/browse/LUCENE-8943 > Project: Lucene - Core > Issue Type: Bug > Components: core/query/scoring >Affects Versions: 8.0 >Reporter: Christoph Goller >Priority: Major > > I recently stumbled across a very old bug in the IDF computation for > MultiPhraseQuery and SpanOrQuery. > BM25Similarity and TFIDFSimilarity / ClassicSimilarity have a method for > combining IDF values from more than on term / TermStatistics. > I mean the method: > Explanation idfExplain(CollectionStatistics collectionStats, TermStatistics > termStats[]) > It simply adds up the IDFs from all termStats[]. > This method is used e.g. in PhraseQuery where it makes sense. If we assume > that for the phrase "New York" the occurrences of both words are independent, > we can multiply their probabilitis and since IDFs are logarithmic we add them > up. Seems to be a reasonable approximation. However, this method is also used > to add up the IDFs of all terms in a MultiPhraseQuery as can be seen in: > Similarity.SimScorer getStats(IndexSearcher searcher) > A MultiPhraseQuery is actually
[jira] [Comment Edited] (LUCENE-8943) Incorrect IDF in MultiPhraseQuery and SpanOrQuery
[ https://issues.apache.org/jira/browse/LUCENE-8943?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16898859#comment-16898859 ] Christoph Goller edited comment on LUCENE-8943 at 8/2/19 12:39 PM: --- Why is this an issue? Because IDFs of SpanOrQueriy and MultiPhraseQuery can get gigantic meaning that such queries have an unexpectedly high impact on the final score. was (Author: gol...@detego-software.de): Why is this an issue? Because IDFs of SpanOrQueriy and MultiPhraseQuery can get gigantic meaning that such queries get an unexpectedly high impact on the final score. > Incorrect IDF in MultiPhraseQuery and SpanOrQuery > - > > Key: LUCENE-8943 > URL: https://issues.apache.org/jira/browse/LUCENE-8943 > Project: Lucene - Core > Issue Type: Bug > Components: core/query/scoring >Affects Versions: 8.0 >Reporter: Christoph Goller >Priority: Major > > I recently stumbled across a very old bug in the IDF computation for > MultiPhraseQuery and SpanOrQuery. > BM25Similarity and TFIDFSimilarity / ClassicSimilarity have a method for > combining IDF values from more than on term / TermStatistics. > I mean the method: > Explanation idfExplain(CollectionStatistics collectionStats, TermStatistics > termStats[]) > It simply adds up the IDFs from all termStats[]. > This method is used e.g. in PhraseQuery where it makes sense. If we assume > that for the phrase "New York" the occurrences of both words are independent, > we can multiply their probabilitis and since IDFs are logarithmic we add them > up. Seems to be a reasonable approximation. However, this method is also used > to add up the IDFs of all terms in a MultiPhraseQuery as can be seen in: > Similarity.SimScorer getStats(IndexSearcher searcher) > A MultiPhraseQuery is actually a PhraseQuery with alternatives at individual > positions. IDFs of alternative terms for one position should not be added up. > Instead we should use the minimum value as an approcimation because this > corresponds to the docFreq of the most frequent term and we know that this is > a lower bound for the docFreq for this position. > In SpanOrQuerry we have the same problem It uses buildSimWeight(...) from > SpanWeight and adds up all IDFs of all OR-clauses. > If my arguments are not convincing, look at SynonymQuery / SynonymWeight in > the constructor: > SynonymWeight(Query query, IndexSearcher searcher, ScoreMode scoreMode, float > boost) > A SynonymQuery is also a kind of OR-query and it uses the maximum of the > docFreq of all its alternative terms. I think this is how it should be. -- This message was sent by Atlassian JIRA (v7.6.14#76016) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (LUCENE-8943) Incorrect IDF in MultiPhraseQuery and SpanOrQuery
[ https://issues.apache.org/jira/browse/LUCENE-8943?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16898859#comment-16898859 ] Christoph Goller commented on LUCENE-8943: -- Why is this an issue? Because IDFs of SpanOrQueriy and MultiPhraseQuery can get gigantic meaning that such queries get an unexpectedly high impact on the final score. > Incorrect IDF in MultiPhraseQuery and SpanOrQuery > - > > Key: LUCENE-8943 > URL: https://issues.apache.org/jira/browse/LUCENE-8943 > Project: Lucene - Core > Issue Type: Bug > Components: core/query/scoring >Affects Versions: 8.0 >Reporter: Christoph Goller >Priority: Major > > I recently stumbled across a very old bug in the IDF computation for > MultiPhraseQuery and SpanOrQuery. > BM25Similarity and TFIDFSimilarity / ClassicSimilarity have a method for > combining IDF values from more than on term / TermStatistics. > I mean the method: > Explanation idfExplain(CollectionStatistics collectionStats, TermStatistics > termStats[]) > It simply adds up the IDFs from all termStats[]. > This method is used e.g. in PhraseQuery where it makes sense. If we assume > that for the phrase "New York" the occurrences of both words are independent, > we can multiply their probabilitis and since IDFs are logarithmic we add them > up. Seems to be a reasonable approximation. However, this method is also used > to add up the IDFs of all terms in a MultiPhraseQuery as can be seen in: > Similarity.SimScorer getStats(IndexSearcher searcher) > A MultiPhraseQuery is actually a PhraseQuery with alternatives at individual > positions. IDFs of alternative terms for one position should not be added up. > Instead we should use the minimum value as an approcimation because this > corresponds to the docFreq of the most frequent term and we know that this is > a lower bound for the docFreq for this position. > In SpanOrQuerry we have the same problem It uses buildSimWeight(...) from > SpanWeight and adds up all IDFs of all OR-clauses. > If my arguments are not convincing, look at SynonymQuery / SynonymWeight in > the constructor: > SynonymWeight(Query query, IndexSearcher searcher, ScoreMode scoreMode, float > boost) > A SynonymQuery is also a kind of OR-query and it uses the maximum of the > docFreq of all its alternative terms. I think this is how it should be. -- This message was sent by Atlassian JIRA (v7.6.14#76016) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Created] (LUCENE-8943) Incorrect IDF in MultiPhraseQuery and SpanOrQuery
Christoph Goller created LUCENE-8943: Summary: Incorrect IDF in MultiPhraseQuery and SpanOrQuery Key: LUCENE-8943 URL: https://issues.apache.org/jira/browse/LUCENE-8943 Project: Lucene - Core Issue Type: Bug Components: core/query/scoring Affects Versions: 8.0 Reporter: Christoph Goller I recently stumbled across a very old bug in the IDF computation for MultiPhraseQuery and SpanOrQuery. BM25Similarity and TFIDFSimilarity / ClassicSimilarity have a method for combining IDF values from more than on term / TermStatistics. I mean the method: Explanation idfExplain(CollectionStatistics collectionStats, TermStatistics termStats[]) It simply adds up the IDFs from all termStats[]. This method is used e.g. in PhraseQuery where it makes sense. If we assume that for the phrase "New York" the occurrences of both words are independent, we can multiply their probabilitis and since IDFs are logarithmic we add them up. Seems to be a reasonable approximation. However, this method is also used to add up the IDFs of all terms in a MultiPhraseQuery as can be seen in: Similarity.SimScorer getStats(IndexSearcher searcher) A MultiPhraseQuery is actually a PhraseQuery with alternatives at individual positions. IDFs of alternative terms for one position should not be added up. Instead we should use the minimum value as an approcimation because this corresponds to the docFreq of the most frequent term and we know that this is a lower bound for the docFreq for this position. In SpanOrQuerry we have the same problem It uses buildSimWeight(...) from SpanWeight and adds up all IDFs of all OR-clauses. If my arguments are not convincing, look at SynonymQuery / SynonymWeight in the constructor: SynonymWeight(Query query, IndexSearcher searcher, ScoreMode scoreMode, float boost) A SynonymQuery is also a kind of OR-query and it uses the maximum of the docFreq of all its alternative terms. I think this is how it should be. -- This message was sent by Atlassian JIRA (v7.6.14#76016) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Created] (LUCENE-8637) WeightedSpanTermExtractor unnexessarily enforces rewrite for some SpanQueiries
Christoph Goller created LUCENE-8637: Summary: WeightedSpanTermExtractor unnexessarily enforces rewrite for some SpanQueiries Key: LUCENE-8637 URL: https://issues.apache.org/jira/browse/LUCENE-8637 Project: Lucene - Core Issue Type: Bug Components: modules/highlighter Affects Versions: 7.5, 7.3.1, 7.4, 7.6 Reporter: Christoph Goller Method mustRewriteQuery(SpanQuery) returns true for SpanPositionCheckQuery, SpanContainingQuery, SpanWithinQuery, and SpanBoostQuery, however, these queries do not require rewriting. One effect of this is e.g. that UnifiedHighlighter does not work with OffsetSource Postings and switches to Analysis which of course has consequences for performance. I attach a patch for lucene version 7.6.0. I have not checked whether it breaks existing unit tests. -- This message was sent by Atlassian JIRA (v7.6.3#76005) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Comment Edited] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
[ https://issues.apache.org/jira/browse/LUCENE-8000?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16214805#comment-16214805 ] Christoph Goller edited comment on LUCENE-8000 at 10/23/17 8:17 AM: ??As an additional point, advanced use cases often utilize token "stacking" for additional uses as well and these would have further distortions on length.?? That's exactly what we are doing. Therefore using discountOverlaps = false could punish languages with more different word forms. I also prefer discountOverlaps = true. I have an intern (student) working on relevance tuning and benchmarks. I think we can try overwriting {code:java} protected float avgFieldLength(CollectionStatistics collectionStats) {code} and see it it changes anything. We will also have a look into Lucene benchmark module. Thanks for your feedback. was (Author: gol...@detego-software.de): ??As an additional point, advanced use cases often utilize token "stacking" for additional uses as well and these would have further distortions on length. ?? That's exactly what we are doing. Therefore using discountOverlaps = false could punish languages with more different word forms. I also prefer discountOverlaps = true. I have an intern (student) working on relevance tuning and benchmarks. I think we can try overwriting {code:java} protected float avgFieldLength(CollectionStatistics collectionStats) {code} and see it it changes anything. We will also have a look into Lucene benchmark module. Thanks for your feedback. > Document Length Normalization in BM25Similarity correct? > > > Key: LUCENE-8000 > URL: https://issues.apache.org/jira/browse/LUCENE-8000 > Project: Lucene - Core > Issue Type: Bug >Reporter: Christoph Goller >Priority: Minor > > Length of individual documents only counts the number of positions of a > document since discountOverlaps defaults to true. > {code} > @Override > public final long computeNorm(FieldInvertState state) { > final int numTerms = discountOverlaps ? state.getLength() - > state.getNumOverlap() : state.getLength(); > int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); > if (indexCreatedVersionMajor >= 7) { > return SmallFloat.intToByte4(numTerms); > } else { > return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); > } > }} > {code} > Measureing document length this way seems perfectly ok for me. What bothers > me is that > average document length is based on sumTotalTermFreq for a field. As far as I > understand that sums up totalTermFreqs for all terms of a field, therefore > counting positions of terms including those that overlap. > {code} > protected float avgFieldLength(CollectionStatistics collectionStats) { > final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); > if (sumTotalTermFreq <= 0) { > return 1f; // field does not exist, or stat is unsupported > } else { > final long docCount = collectionStats.docCount() == -1 ? > collectionStats.maxDoc() : collectionStats.docCount(); > return (float) (sumTotalTermFreq / (double) docCount); > } > } > } > {code} > Are we comparing apples and oranges in the final scoring? > I haven't run any benchmarks and I am not sure whether this has a serious > effect. It just means that documents that have synonyms or in my use case > different normal forms of tokens on the same position are shorter and > therefore get higher scores than they should and that we do not use the > whole spectrum of relative document lenght of BM25. > I think for BM25 discountOverlaps should default to false. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
[ https://issues.apache.org/jira/browse/LUCENE-8000?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16214805#comment-16214805 ] Christoph Goller commented on LUCENE-8000: -- ??As an additional point, advanced use cases often utilize token "stacking" for additional uses as well and these would have further distortions on length. ?? That's exactly what we are doing. Therefore using discountOverlaps = false could punish languages with more different word forms. I also prefer discountOverlaps = true. I have an intern (student) working on relevance tuning and benchmarks. I think we can try overwriting {code:java} protected float avgFieldLength(CollectionStatistics collectionStats) {code} and see it it changes anything. We will also have a look into Lucene benchmark module. Thanks for your feedback. > Document Length Normalization in BM25Similarity correct? > > > Key: LUCENE-8000 > URL: https://issues.apache.org/jira/browse/LUCENE-8000 > Project: Lucene - Core > Issue Type: Bug >Reporter: Christoph Goller >Priority: Minor > > Length of individual documents only counts the number of positions of a > document since discountOverlaps defaults to true. > {code} > @Override > public final long computeNorm(FieldInvertState state) { > final int numTerms = discountOverlaps ? state.getLength() - > state.getNumOverlap() : state.getLength(); > int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); > if (indexCreatedVersionMajor >= 7) { > return SmallFloat.intToByte4(numTerms); > } else { > return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); > } > }} > {code} > Measureing document length this way seems perfectly ok for me. What bothers > me is that > average document length is based on sumTotalTermFreq for a field. As far as I > understand that sums up totalTermFreqs for all terms of a field, therefore > counting positions of terms including those that overlap. > {code} > protected float avgFieldLength(CollectionStatistics collectionStats) { > final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); > if (sumTotalTermFreq <= 0) { > return 1f; // field does not exist, or stat is unsupported > } else { > final long docCount = collectionStats.docCount() == -1 ? > collectionStats.maxDoc() : collectionStats.docCount(); > return (float) (sumTotalTermFreq / (double) docCount); > } > } > } > {code} > Are we comparing apples and oranges in the final scoring? > I haven't run any benchmarks and I am not sure whether this has a serious > effect. It just means that documents that have synonyms or in my use case > different normal forms of tokens on the same position are shorter and > therefore get higher scores than they should and that we do not use the > whole spectrum of relative document lenght of BM25. > I think for BM25 discountOverlaps should default to false. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Comment Edited] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
[ https://issues.apache.org/jira/browse/LUCENE-8000?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16212350#comment-16212350 ] Christoph Goller edited comment on LUCENE-8000 at 10/20/17 8:42 AM: ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood. *I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explanation why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explanation that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Another possible explanation would be that some fields have synonyms and others have not. That would punish fields with synonyms compared to others since their length is greater (in Classic Similarity with discountOverlaps = false), but in BM25 it should not have this effect since BM25 uses relative lenght for scoring and not abolute length like Classic Similarity. Sorry for bothering you with these questions. It's only my curiosity and maybe Jira is not the right place for this. was (Author: gol...@detego-software.de): ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood. *I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explanation that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Another possible explanation would be that some fields have synonyms and others have not. That would punish fields with synonyms compared to others since their length is greater (in Classic Similarity with discountOverlaps = false), but in BM25 it should not have this effect since BM25 uses relative lenght for scoring and not abolute length like Classic Similarity. Sorry for bothering you with these questions. It's only my curiosity and maybe Jira is not the right place for this. > Document Length Normalization in BM25Similarity correct? > > > Key: LUCENE-8000 > URL: https://issues.apache.org/jira/browse/LUCENE-8000 > Project: Lucene - Core > Issue Type: Bug >Reporter: Christoph Goller >Priority: Minor > > Length of individual documents only counts the number of positions of a > document since discountOverlaps defaults to true. > {code} > @Override > public final long computeNorm(FieldInvertState state) { > final int numTerms = discountOverlaps ? state.getLength() - > state.getNumOverlap() : state.getLength(); > int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); > if (indexCreatedVersionMajor >= 7) { > return SmallFloat.intToByte4(numTerms); > } else { > return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); > } > }} > {code} > Measureing document length this way seems perfectly ok for me. What bothers > me is that > average document length is based on sumTotalTermFreq for a field. As far as I > understand that sums up totalTermFreqs for all terms of a field, therefore > counting positions of terms includin
[jira] [Comment Edited] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
[ https://issues.apache.org/jira/browse/LUCENE-8000?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16212350#comment-16212350 ] Christoph Goller edited comment on LUCENE-8000 at 10/20/17 8:42 AM: ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood. *I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explanation that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Another possible explanation would be that some fields have synonyms and others have not. That would punish fields with synonyms compared to others since their length is greater (in Classic Similarity with discountOverlaps = false), but in BM25 it should not have this effect since BM25 uses relative lenght for scoring and not abolute length like Classic Similarity. Sorry for bothering you with these questions. It's only my curiosity and maybe Jira is not the right place for this. was (Author: gol...@detego-software.de): ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood. *I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explanation that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Another possible explanation would be that some fields have synonyms and others have not. That would punish fields with synonyms compared to others since their length is greater (in Classic Similarity with discountOverlaps = false), but in BM25 it should not have this effect since BM25 uses relative lenght for scoring and not abolute length like Classic Similarity. Sorry for bothering you with these questions. It's only my curiosity and maybe Jira is nto the right place for this. > Document Length Normalization in BM25Similarity correct? > > > Key: LUCENE-8000 > URL: https://issues.apache.org/jira/browse/LUCENE-8000 > Project: Lucene - Core > Issue Type: Bug >Reporter: Christoph Goller >Priority: Minor > > Length of individual documents only counts the number of positions of a > document since discountOverlaps defaults to true. > {code} > @Override > public final long computeNorm(FieldInvertState state) { > final int numTerms = discountOverlaps ? state.getLength() - > state.getNumOverlap() : state.getLength(); > int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); > if (indexCreatedVersionMajor >= 7) { > return SmallFloat.intToByte4(numTerms); > } else { > return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); > } > }} > {code} > Measureing document length this way seems perfectly ok for me. What bothers > me is that > average document length is based on sumTotalTermFreq for a field. As far as I > understand that sums up totalTermFreqs for all terms of a field, therefore > counting positions of terms includi
[jira] [Comment Edited] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
[ https://issues.apache.org/jira/browse/LUCENE-8000?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16212350#comment-16212350 ] Christoph Goller edited comment on LUCENE-8000 at 10/20/17 8:41 AM: ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood. *I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explanation that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Another possible explanation would be that some fields have synonyms and others have not. That would punish fields with synonyms compared to others since their length is greater (in Classic Similarity with discountOverlaps = false), but in BM25 it should not have this effect since BM25 uses relative lenght for scoring and not abolute length like Classic Similarity. Sorry for bothering you with these questions. It's only my curiosity and maybe Jira is nto the right place for this. was (Author: gol...@detego-software.de): ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood. *I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explanation that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Another possible explanation would be that some fields have synonyms and others have not. That would punish fields with synonyms compared to others since their length is greater (in Classic Similarity with discountOverlaps = false), but in BM25 it should not have this effect since BM25 used relative lenght for scoring and not abolute length. Sorry for bothering you with these questions. It's only my curiosity and maybe Jira is nto the right place for this. > Document Length Normalization in BM25Similarity correct? > > > Key: LUCENE-8000 > URL: https://issues.apache.org/jira/browse/LUCENE-8000 > Project: Lucene - Core > Issue Type: Bug >Reporter: Christoph Goller >Priority: Minor > > Length of individual documents only counts the number of positions of a > document since discountOverlaps defaults to true. > {code} > @Override > public final long computeNorm(FieldInvertState state) { > final int numTerms = discountOverlaps ? state.getLength() - > state.getNumOverlap() : state.getLength(); > int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); > if (indexCreatedVersionMajor >= 7) { > return SmallFloat.intToByte4(numTerms); > } else { > return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); > } > }} > {code} > Measureing document length this way seems perfectly ok for me. What bothers > me is that > average document length is based on sumTotalTermFreq for a field. As far as I > understand that sums up totalTermFreqs for all terms of a field, therefore > counting positions of terms including those that overlap. >
[jira] [Comment Edited] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
[ https://issues.apache.org/jira/browse/LUCENE-8000?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16212350#comment-16212350 ] Christoph Goller edited comment on LUCENE-8000 at 10/20/17 8:39 AM: ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood. *I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explanation that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Another possible explanation would be that some fields have synonyms and others have not. That would punish fields with synonyms compared to others since their length is greater (in Classic Similarity with discountOverlaps = false), but in BM25 it should not have this effect since BM25 used relative lenght for scoring and not abolute length. Sorry for bothering you with these questions. It's only my curiosity and maybe Jira is nto the right place for this. was (Author: gol...@detego-software.de): ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood. *I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explanation that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Sorry for bothering you with these questions. It's only my curiosity and maybe Jira is nto the right place for this. > Document Length Normalization in BM25Similarity correct? > > > Key: LUCENE-8000 > URL: https://issues.apache.org/jira/browse/LUCENE-8000 > Project: Lucene - Core > Issue Type: Bug >Reporter: Christoph Goller >Priority: Minor > > Length of individual documents only counts the number of positions of a > document since discountOverlaps defaults to true. > {code} > @Override > public final long computeNorm(FieldInvertState state) { > final int numTerms = discountOverlaps ? state.getLength() - > state.getNumOverlap() : state.getLength(); > int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); > if (indexCreatedVersionMajor >= 7) { > return SmallFloat.intToByte4(numTerms); > } else { > return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); > } > }} > {code} > Measureing document length this way seems perfectly ok for me. What bothers > me is that > average document length is based on sumTotalTermFreq for a field. As far as I > understand that sums up totalTermFreqs for all terms of a field, therefore > counting positions of terms including those that overlap. > {code} > protected float avgFieldLength(CollectionStatistics collectionStats) { > final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); > if (sumTotalTermFreq <= 0) { > return 1f; // field does not exist, or stat is unsupported > } else { > final long docCount = collectionStats.docCount() == -1 ? > collectionStats.maxDoc() : co
[jira] [Comment Edited] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
[ https://issues.apache.org/jira/browse/LUCENE-8000?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16212350#comment-16212350 ] Christoph Goller edited comment on LUCENE-8000 at 10/20/17 8:35 AM: ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood. *I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explanation that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Sorry for bothering you with these questions. It's only my curiosity and maybe Jira is nto the right place for this. was (Author: gol...@detego-software.de): ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood. *I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explaination that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Sorry for bothering you with these questions. It's only my curiosity and maybe Jira is nto the right place for this. > Document Length Normalization in BM25Similarity correct? > > > Key: LUCENE-8000 > URL: https://issues.apache.org/jira/browse/LUCENE-8000 > Project: Lucene - Core > Issue Type: Bug >Reporter: Christoph Goller >Priority: Minor > > Length of individual documents only counts the number of positions of a > document since discountOverlaps defaults to true. > {code} > @Override > public final long computeNorm(FieldInvertState state) { > final int numTerms = discountOverlaps ? state.getLength() - > state.getNumOverlap() : state.getLength(); > int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); > if (indexCreatedVersionMajor >= 7) { > return SmallFloat.intToByte4(numTerms); > } else { > return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); > } > }} > {code} > Measureing document length this way seems perfectly ok for me. What bothers > me is that > average document length is based on sumTotalTermFreq for a field. As far as I > understand that sums up totalTermFreqs for all terms of a field, therefore > counting positions of terms including those that overlap. > {code} > protected float avgFieldLength(CollectionStatistics collectionStats) { > final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); > if (sumTotalTermFreq <= 0) { > return 1f; // field does not exist, or stat is unsupported > } else { > final long docCount = collectionStats.docCount() == -1 ? > collectionStats.maxDoc() : collectionStats.docCount(); > return (float) (sumTotalTermFreq / (double) docCount); > } > } > } > {code} > Are we comparing apples and oranges in the final scoring? > I haven't run any benchmarks and I am not sure whether this has a serious > effect. It just means that documents that have synonyms or in my use case > different norma
[jira] [Commented] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
[ https://issues.apache.org/jira/browse/LUCENE-8000?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16212350#comment-16212350 ] Christoph Goller commented on LUCENE-8000: -- ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood.* I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explaination that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Sorry for bothering you with these questions. It's only my curiosity and mayb Jira is nto the right place for this. > Document Length Normalization in BM25Similarity correct? > > > Key: LUCENE-8000 > URL: https://issues.apache.org/jira/browse/LUCENE-8000 > Project: Lucene - Core > Issue Type: Bug >Reporter: Christoph Goller >Priority: Minor > > Length of individual documents only counts the number of positions of a > document since discountOverlaps defaults to true. > {code} > @Override > public final long computeNorm(FieldInvertState state) { > final int numTerms = discountOverlaps ? state.getLength() - > state.getNumOverlap() : state.getLength(); > int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); > if (indexCreatedVersionMajor >= 7) { > return SmallFloat.intToByte4(numTerms); > } else { > return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); > } > }} > {code} > Measureing document length this way seems perfectly ok for me. What bothers > me is that > average document length is based on sumTotalTermFreq for a field. As far as I > understand that sums up totalTermFreqs for all terms of a field, therefore > counting positions of terms including those that overlap. > {code} > protected float avgFieldLength(CollectionStatistics collectionStats) { > final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); > if (sumTotalTermFreq <= 0) { > return 1f; // field does not exist, or stat is unsupported > } else { > final long docCount = collectionStats.docCount() == -1 ? > collectionStats.maxDoc() : collectionStats.docCount(); > return (float) (sumTotalTermFreq / (double) docCount); > } > } > } > {code} > Are we comparing apples and oranges in the final scoring? > I haven't run any benchmarks and I am not sure whether this has a serious > effect. It just means that documents that have synonyms or in my use case > different normal forms of tokens on the same position are shorter and > therefore get higher scores than they should and that we do not use the > whole spectrum of relative document lenght of BM25. > I think for BM25 discountOverlaps should default to false. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Comment Edited] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
[ https://issues.apache.org/jira/browse/LUCENE-8000?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16212350#comment-16212350 ] Christoph Goller edited comment on LUCENE-8000 at 10/20/17 8:34 AM: ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood. *I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explaination that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Sorry for bothering you with these questions. It's only my curiosity and maybe Jira is nto the right place for this. was (Author: gol...@detego-software.de): ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood. *I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explaination that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Sorry for bothering you with these questions. It's only my curiosity and mayb Jira is nto the right place for this. > Document Length Normalization in BM25Similarity correct? > > > Key: LUCENE-8000 > URL: https://issues.apache.org/jira/browse/LUCENE-8000 > Project: Lucene - Core > Issue Type: Bug >Reporter: Christoph Goller >Priority: Minor > > Length of individual documents only counts the number of positions of a > document since discountOverlaps defaults to true. > {code} > @Override > public final long computeNorm(FieldInvertState state) { > final int numTerms = discountOverlaps ? state.getLength() - > state.getNumOverlap() : state.getLength(); > int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); > if (indexCreatedVersionMajor >= 7) { > return SmallFloat.intToByte4(numTerms); > } else { > return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); > } > }} > {code} > Measureing document length this way seems perfectly ok for me. What bothers > me is that > average document length is based on sumTotalTermFreq for a field. As far as I > understand that sums up totalTermFreqs for all terms of a field, therefore > counting positions of terms including those that overlap. > {code} > protected float avgFieldLength(CollectionStatistics collectionStats) { > final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); > if (sumTotalTermFreq <= 0) { > return 1f; // field does not exist, or stat is unsupported > } else { > final long docCount = collectionStats.docCount() == -1 ? > collectionStats.maxDoc() : collectionStats.docCount(); > return (float) (sumTotalTermFreq / (double) docCount); > } > } > } > {code} > Are we comparing apples and oranges in the final scoring? > I haven't run any benchmarks and I am not sure whether this has a serious > effect. It just means that documents that have synonyms or in my use case > different norma
[jira] [Comment Edited] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
[ https://issues.apache.org/jira/browse/LUCENE-8000?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16212350#comment-16212350 ] Christoph Goller edited comment on LUCENE-8000 at 10/20/17 8:34 AM: ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood. *I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explaination that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Sorry for bothering you with these questions. It's only my curiosity and mayb Jira is nto the right place for this. was (Author: gol...@detego-software.de): ??My point is that defaults are for typical use-cases, and the default of discountOverlaps meets that goal. It results in better (measured) performance for many tokenfilters that are commonly used such as common-grams, WDF, synonyms, etc. I ran these tests before proposing the default, it was not done flying blind.?? Understood.* I have not experienced any problems with the current default* and I have the option to set discountOverlaps to false. Therefore it's ok for me if the ticket gets closed. I only think about this out of "scientific" curiosity in the context of relevance tuning. What benchmarks have you used for measuring performance? Is your opinion based on tests with Lucene Classic Similarity (it also uses discountOverlaps = true) or also on tests with BM25. Have you any idea / explaination why relevancy is better using discountOverlaps = true. My naive guess would be that since stopwords or synonyms are either used on all documents or on none and therefore it should not make much difference whether we count overlaps or not. Is the explaination that for some documents many stopwords / synonyms / WDF splits are used and for others not (for the same field). Sorry for bothering you with these questions. It's only my curiosity and mayb Jira is nto the right place for this. > Document Length Normalization in BM25Similarity correct? > > > Key: LUCENE-8000 > URL: https://issues.apache.org/jira/browse/LUCENE-8000 > Project: Lucene - Core > Issue Type: Bug >Reporter: Christoph Goller >Priority: Minor > > Length of individual documents only counts the number of positions of a > document since discountOverlaps defaults to true. > {code} > @Override > public final long computeNorm(FieldInvertState state) { > final int numTerms = discountOverlaps ? state.getLength() - > state.getNumOverlap() : state.getLength(); > int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); > if (indexCreatedVersionMajor >= 7) { > return SmallFloat.intToByte4(numTerms); > } else { > return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); > } > }} > {code} > Measureing document length this way seems perfectly ok for me. What bothers > me is that > average document length is based on sumTotalTermFreq for a field. As far as I > understand that sums up totalTermFreqs for all terms of a field, therefore > counting positions of terms including those that overlap. > {code} > protected float avgFieldLength(CollectionStatistics collectionStats) { > final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); > if (sumTotalTermFreq <= 0) { > return 1f; // field does not exist, or stat is unsupported > } else { > final long docCount = collectionStats.docCount() == -1 ? > collectionStats.maxDoc() : collectionStats.docCount(); > return (float) (sumTotalTermFreq / (double) docCount); > } > } > } > {code} > Are we comparing apples and oranges in the final scoring? > I haven't run any benchmarks and I am not sure whether this has a serious > effect. It just means that documents that have synonyms or in my use case > different normal
[jira] [Updated] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
[ https://issues.apache.org/jira/browse/LUCENE-8000?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Christoph Goller updated LUCENE-8000: - Description: Length of individual documents only counts the number of positions of a document since discountOverlaps defaults to true. {code} @Override public final long computeNorm(FieldInvertState state) { final int numTerms = discountOverlaps ? state.getLength() - state.getNumOverlap() : state.getLength(); int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); if (indexCreatedVersionMajor >= 7) { return SmallFloat.intToByte4(numTerms); } else { return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); } }} {code} Measureing document length this way seems perfectly ok for me. What bothers me is that average document length is based on sumTotalTermFreq for a field. As far as I understand that sums up totalTermFreqs for all terms of a field, therefore counting positions of terms including those that overlap. {code} protected float avgFieldLength(CollectionStatistics collectionStats) { final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); if (sumTotalTermFreq <= 0) { return 1f; // field does not exist, or stat is unsupported } else { final long docCount = collectionStats.docCount() == -1 ? collectionStats.maxDoc() : collectionStats.docCount(); return (float) (sumTotalTermFreq / (double) docCount); } } } {code} Are we comparing apples and oranges in the final scoring? I haven't run any benchmarks and I am not sure whether this has a serious effect. It just means that documents that have synonyms or in my use case different normal forms of tokens on the same position are shorter and therefore get higher scores than they should and that we do not use the whole spectrum of relative document lenght of BM25. I think for BM25 discountOverlaps should default to false. was: Length of individual documents only counts the number of positions of a document since discountOverlaps defaults to true. {code} @Override public final long computeNorm(FieldInvertState state) { final int numTerms = discountOverlaps ? state.getLength() - state.getNumOverlap() : state.getLength(); int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); if (indexCreatedVersionMajor >= 7) { return SmallFloat.intToByte4(numTerms); } else { return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); } }} {code} Measureing document length this way seems perfectly ok for me. What bothers me is that average document length is based on sumTotalTermFreq for a field. As far as I understand that sums up totalTermFreqs for all terms of a field, therefore counting positions of terms including those that overlap. {code} protected float avgFieldLength(CollectionStatistics collectionStats) { final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); if (sumTotalTermFreq <= 0) { return 1f; // field does not exist, or stat is unsupported } else { final long docCount = collectionStats.docCount() == -1 ? collectionStats.maxDoc() : collectionStats.docCount(); return (float) (sumTotalTermFreq / (double) docCount); } } } {code} Are we comparing apples and oranges in the final scoring? I haven't run any benchmarks and I am not sure whether this has a serious effect. It just means that documents that have synonyms or in our case different normal forms of tokens on the same position are shorter and therefore get higher scores than they should and that we do not use the whole spectrum of relative document lenght of BM25. I think for BM25 discountOverlaps should default to false. > Document Length Normalization in BM25Similarity correct? > > > Key: LUCENE-8000 > URL: https://issues.apache.org/jira/browse/LUCENE-8000 > Project: Lucene - Core > Issue Type: Bug >Reporter: Christoph Goller >Priority: Minor > > Length of individual documents only counts the number of positions of a > document since discountOverlaps defaults to true. > {code} > @Override > public final long computeNorm(FieldInvertState state) { > final int numTerms = discountOverlaps ? state.getLength() - > state.getNumOverlap() : state.getLength(); > int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); > if (indexCreatedVersionMajor >= 7) { > return SmallFloat.intToByte4(numTerms); > } else { > return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); > } > }} > {code} > Measureing document length this way seems perfectly ok for me. What bothers > me is that > average document length is based on sumTotalTermFreq for a field. As far as I > und
[jira] [Updated] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
[ https://issues.apache.org/jira/browse/LUCENE-8000?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Christoph Goller updated LUCENE-8000: - Description: Length of individual documents only counts the number of positions of a document since discountOverlaps defaults to true. {code} @Override public final long computeNorm(FieldInvertState state) { final int numTerms = discountOverlaps ? state.getLength() - state.getNumOverlap() : state.getLength(); int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); if (indexCreatedVersionMajor >= 7) { return SmallFloat.intToByte4(numTerms); } else { return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); } }} {code} Measureing document length this way seems perfectly ok for me. What bothers me is that average document length is based on sumTotalTermFreq for a field. As far as I understand that sums up totalTermFreqs for all terms of a field, therefore counting positions of terms including those that overlap. {code} protected float avgFieldLength(CollectionStatistics collectionStats) { final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); if (sumTotalTermFreq <= 0) { return 1f; // field does not exist, or stat is unsupported } else { final long docCount = collectionStats.docCount() == -1 ? collectionStats.maxDoc() : collectionStats.docCount(); return (float) (sumTotalTermFreq / (double) docCount); } } } {code} Are we comparing apples and oranges in the final scoring? I haven't run any benchmarks and I am not sure whether this has a serious effect. It just means that documents that have synonyms or in our case different normal forms of tokens on the same position are shorter and therefore get higher scores than they should and that we do not use the whole spectrum of relative document lenght of BM25. I think for BM25 discountOverlaps should default to false. was: Length of individual documents only counts the number of positions of a document since discountOverlaps defaults to true. { @Override public final long computeNorm(FieldInvertState state) { final int numTerms = discountOverlaps ? state.getLength() - state.getNumOverlap() : state.getLength(); int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); if (indexCreatedVersionMajor >= 7) { return SmallFloat.intToByte4(numTerms); } else { return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); } }}} Measureing document length this way seems perfectly ok for me. What bothers me is that average document length is based on sumTotalTermFreq for a field. As far as I understand that sums up totalTermFreqs for all terms of a field, therefore counting positions of terms including those that overlap. {{ protected float avgFieldLength(CollectionStatistics collectionStats) { final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); if (sumTotalTermFreq <= 0) { return 1f; // field does not exist, or stat is unsupported } else { final long docCount = collectionStats.docCount() == -1 ? collectionStats.maxDoc() : collectionStats.docCount(); return (float) (sumTotalTermFreq / (double) docCount); } } }} Are we comparing apples and oranges in the final scoring? I haven't run any benchmarks and I am not sure whether this has a serious effect. It just means that documents that have synonyms or in our case different normal forms of tokens on the same position are shorter and therefore get higher scores than they should and that we do not use the whole spectrum of relative document lenght of BM25. I think for BM25 discountOverlaps should default to false. > Document Length Normalization in BM25Similarity correct? > > > Key: LUCENE-8000 > URL: https://issues.apache.org/jira/browse/LUCENE-8000 > Project: Lucene - Core > Issue Type: Bug >Reporter: Christoph Goller >Priority: Minor > > Length of individual documents only counts the number of positions of a > document since discountOverlaps defaults to true. > {code} > @Override > public final long computeNorm(FieldInvertState state) { > final int numTerms = discountOverlaps ? state.getLength() - > state.getNumOverlap() : state.getLength(); > int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); > if (indexCreatedVersionMajor >= 7) { > return SmallFloat.intToByte4(numTerms); > } else { > return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); > } > }} > {code} > Measureing document length this way seems perfectly ok for me. What bothers > me is that > average document length is based on sumTotalTermFreq for a field. As far as I > understand that sums up tota
[jira] [Updated] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
[ https://issues.apache.org/jira/browse/LUCENE-8000?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Christoph Goller updated LUCENE-8000: - Description: Length of individual documents only counts the number of positions of a document since discountOverlaps defaults to true. { @Override public final long computeNorm(FieldInvertState state) { final int numTerms = discountOverlaps ? state.getLength() - state.getNumOverlap() : state.getLength(); int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); if (indexCreatedVersionMajor >= 7) { return SmallFloat.intToByte4(numTerms); } else { return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); } }}} Measureing document length this way seems perfectly ok for me. What bothers me is that average document length is based on sumTotalTermFreq for a field. As far as I understand that sums up totalTermFreqs for all terms of a field, therefore counting positions of terms including those that overlap. {{ protected float avgFieldLength(CollectionStatistics collectionStats) { final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); if (sumTotalTermFreq <= 0) { return 1f; // field does not exist, or stat is unsupported } else { final long docCount = collectionStats.docCount() == -1 ? collectionStats.maxDoc() : collectionStats.docCount(); return (float) (sumTotalTermFreq / (double) docCount); } } }} Are we comparing apples and oranges in the final scoring? I haven't run any benchmarks and I am not sure whether this has a serious effect. It just means that documents that have synonyms or in our case different normal forms of tokens on the same position are shorter and therefore get higher scores than they should and that we do not use the whole spectrum of relative document lenght of BM25. I think for BM25 discountOverlaps should default to false. was: Length of individual documents only counts the number of positions of a document since discountOverlaps defaults to true. {quote} @Override public final long computeNorm(FieldInvertState state) { final int numTerms = discountOverlaps ? state.getLength() - state.getNumOverlap() : state.getLength(); int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); if (indexCreatedVersionMajor >= 7) { return SmallFloat.intToByte4(numTerms); } else { return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); } }{quote} Measureing document length this way seems perfectly ok for me. What bothers me is that average document length is based on sumTotalTermFreq for a field. As far as I understand that sums up totalTermFreqs for all terms of a field, therefore counting positions of terms including those that overlap. {quote} protected float avgFieldLength(CollectionStatistics collectionStats) { final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); if (sumTotalTermFreq <= 0) { return 1f; // field does not exist, or stat is unsupported } else { final long docCount = collectionStats.docCount() == -1 ? collectionStats.maxDoc() : collectionStats.docCount(); return (float) (sumTotalTermFreq / (double) docCount); } }{quote} Are we comparing apples and oranges in the final scoring? I haven't run any benchmarks and I am not sure whether this has a serious effect. It just means that documents that have synonyms or in our case different normal forms of tokens on the same position are shorter and therefore get higher scores than they should and that we do not use the whole spectrum of relative document lenght of BM25. I think for BM25 discountOverlaps should default to false. > Document Length Normalization in BM25Similarity correct? > > > Key: LUCENE-8000 > URL: https://issues.apache.org/jira/browse/LUCENE-8000 > Project: Lucene - Core > Issue Type: Bug >Reporter: Christoph Goller >Priority: Minor > > Length of individual documents only counts the number of positions of a > document since discountOverlaps defaults to true. > { @Override > public final long computeNorm(FieldInvertState state) { > final int numTerms = discountOverlaps ? state.getLength() - > state.getNumOverlap() : state.getLength(); > int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); > if (indexCreatedVersionMajor >= 7) { > return SmallFloat.intToByte4(numTerms); > } else { > return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); > } > }}} > Measureing document length this way seems perfectly ok for me. What bothers > me is that > average document length is based on sumTotalTermFreq for a field. As far as I > understand that sums up totalTermFreqs for a
[jira] [Created] (LUCENE-8000) Document Length Normalization in BM25Similarity correct?
Christoph Goller created LUCENE-8000: Summary: Document Length Normalization in BM25Similarity correct? Key: LUCENE-8000 URL: https://issues.apache.org/jira/browse/LUCENE-8000 Project: Lucene - Core Issue Type: Bug Reporter: Christoph Goller Priority: Minor Length of individual documents only counts the number of positions of a document since discountOverlaps defaults to true. {quote} @Override public final long computeNorm(FieldInvertState state) { final int numTerms = discountOverlaps ? state.getLength() - state.getNumOverlap() : state.getLength(); int indexCreatedVersionMajor = state.getIndexCreatedVersionMajor(); if (indexCreatedVersionMajor >= 7) { return SmallFloat.intToByte4(numTerms); } else { return SmallFloat.floatToByte315((float) (1 / Math.sqrt(numTerms))); } }{quote} Measureing document length this way seems perfectly ok for me. What bothers me is that average document length is based on sumTotalTermFreq for a field. As far as I understand that sums up totalTermFreqs for all terms of a field, therefore counting positions of terms including those that overlap. {quote} protected float avgFieldLength(CollectionStatistics collectionStats) { final long sumTotalTermFreq = collectionStats.sumTotalTermFreq(); if (sumTotalTermFreq <= 0) { return 1f; // field does not exist, or stat is unsupported } else { final long docCount = collectionStats.docCount() == -1 ? collectionStats.maxDoc() : collectionStats.docCount(); return (float) (sumTotalTermFreq / (double) docCount); } }{quote} Are we comparing apples and oranges in the final scoring? I haven't run any benchmarks and I am not sure whether this has a serious effect. It just means that documents that have synonyms or in our case different normal forms of tokens on the same position are shorter and therefore get higher scores than they should and that we do not use the whole spectrum of relative document lenght of BM25. I think for BM25 discountOverlaps should default to false. -- This message was sent by Atlassian JIRA (v6.4.14#64029) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15470053#comment-15470053 ] Christoph Goller commented on LUCENE-7398: -- I just found that the LUCENE-2878 work/branch may contain some interesting ideas about scoring and proximity search / Span*Queries. > Nested Span Queries are buggy > - > > Key: LUCENE-7398 > URL: https://issues.apache.org/jira/browse/LUCENE-7398 > Project: Lucene - Core > Issue Type: Bug > Components: core/search >Affects Versions: 5.5, 6.x >Reporter: Christoph Goller >Assignee: Alan Woodward >Priority: Critical > Attachments: LUCENE-7398-20160814.patch, LUCENE-7398.patch, > LUCENE-7398.patch, TestSpanCollection.java > > > Example for a nested SpanQuery that is not working: > Document: Human Genome Organization , HUGO , is trying to coordinate gene > mapping research worldwide. > Query: spanNear([body:coordinate, spanOr([spanNear([body:gene, body:mapping], > 0, true), body:gene]), body:research], 0, true) > The query should match "coordinate gene mapping research" as well as > "coordinate gene research". It does not match "coordinate gene mapping > research" with Lucene 5.5 or 6.1, it did however match with Lucene 4.10.4. It > probably stopped working with the changes on SpanQueries in 5.3. I will > attach a unit test that shows the problem. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Comment Edited] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15466888#comment-15466888 ] Christoph Goller edited comment on LUCENE-7398 at 9/6/16 3:37 PM: -- After thoroughly reviewing the current implementations of SpanNearQuery, PhraseQuery and MultiPhraseQuery I see some problems and inconsistencies. I volunteer to fix at least some of these problems, but first I would like to have a consensus about the desired bahavior of SpanQuery. This ticket may not be the right place for such a discussion, so please point me to a better place if there is one. 1) Missing Matches caused by lazy iteration: I think lazy iteration is not a new thing in Lucene SpanNearQuery. As far as I know there never was an implementation that compared all possible combinations of subspan matches for SpanNearQuery in Lucene. So SpanNearQuery always missed some matches. *) This ticket demonstrates missing matches for ordered SpanQuery. Documents that should match don't match. This is caused by subspans of SpanNearQuery having a variable match length. For these cases the lazy iteration implementation which tries to optimize the number of comparisons of subspan matches is not sufficient. *) Tim tried these examples with unorderd SpanQuery and got the same bahavior. I think this is caused by a similar kind of lazy iteration in the unordered case. *) In the unordered case lazy iteration also causes problems if the subspans do not have variable-length matches. This is demonstrated in LUCENE-5331 and LUCENE-2861. Tim, thanks for pointing to these tickets. In these examples all clauses of the SpanNearQuery were SpanTermQueries, but some occured more than once. For PhraseQuery and MultiPhraseQuery and their implementation in SloppyPhraseScore this seems to be a known problem that has been solved by a special complex treatment of repetitions that I currently don't understand in detail. My current opinion: We should give up lazy iteration for the unordered and the ordered case to solve these problems. I think it can be done and the performance peanalty should not be too big. We already iterate over all positions of all subspans. So we already have done the expensive operation of reading them. Should some more comparisons of int-values (positions) really matter so much? At least for the ordered case I am optimistic that I could implement it efficiently. 2) Inconsistent Scoring of SpanNearQuery *) Lazy iteration means that some "redundant" matches in a document are skipped in order to have a faster matching algorithm. I am not sure how redundant was defined exactly for the idea of lazy iteration. It referred to matches with the same start posisiton somehow. As long as different matches for the first clause are concerned, they are found, but not the all matches for intermediate subclauses are regarded. Skipping matches however reduces the frequency that is computed and consequently the score. See Javadoc of phraseFreq() in SloppyPhraseScore which mention the same phenomenon. This is quite important for my use case of SpanQueries. I have different versions/variants of the same term on the same position, e.g. one with case-normalization and one without and I want a higher score if the user-query matches for more than one variant, and I use this approach for clauses of SpanNearQuery. *) In NearSpansOrdered the method width() (it is used to compute sloppy frequency in SpanScore) returns the number of gaps between the matches. If you have a perfect match it returns 0 (no sloppyness). In NearSpansUnordered it returns the length of the match, not the number of gaps. See atMatch() for the difference. The reason is probably, that (maxEndPositionCell.endPosition() - minPositionCell().startPosition() - totalSpanLength) might even become negative if matches overlap. I would prefer something like Math.max(0, (maxEndPositionCell.endPosition() - minPositionCell().startPosition() - totalSpanLength)) *) SpanOrQuery and SpanNearQuery completely ignore the scores of their subclauses (subweights are always generated as non-scoring). A SpanOrQuery should give a Score similar to a BooleanQuery, shouldn't it? As long as we have this behavior, SpanBoostQuery does not make any sense, doese it? So to my opinion the existance of SpanBoostQuery shows that others also had the idea that a nested SpanQuery should somehow use the scores of their clauses for the computation of their own score. was (Author: gol...@detego-software.de): After thoroughly reviewing the current implementations of SpanNearQuery, PhraseQuery and MultiPhraseQuery I see some problems and inconsistencies. I volunteer to fix at least some of these problems, but first I would like to have a consensus about the desired bahavior of SpanQuery. This ticket may not be the right place for such a d
[jira] [Comment Edited] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15466888#comment-15466888 ] Christoph Goller edited comment on LUCENE-7398 at 9/6/16 9:36 AM: -- After thoroughly reviewing the current implementations of SpanNearQuery, PhraseQuery and MultiPhraseQuery I see some problems and inconsistencies. I volunteer to fix at least some of these problems, but first I would like to have a consensus about the desired bahavior of SpanQuery. This ticket may not be the right place for such a discussion, so please point me to a better place if there is one. 1) Missing Matches caused by lazy iteration: I think lazy iteration is not a new thing in Lucene SpanNearQuery. As far as I know there never was an implementation that compared all possible combinations of subspan matches for SpanNearQuery in Lucene. So SpanNEarQuery always missed some matches. *) This ticket demonstrates missing matches for ordered SpanQuery. Documents that should match aer not matching. They are caused by subspans of SpanNearQuery having a variable match length. For these cases the lazy iteration implementation which tries to optimize the number of comparisons of subspan matches is not sufficient. *) Tim tried these examples with unorderd SpanQuery and got the same bahavior. I think this is caused by a similar kind of lazy iteration in the unordered case. *) In the unorderd case lazy iteration also causes problems if the subspans do not have variable-length matches. This is demonstrated in LUCENE-5331 and LUCENE-2861. Tim, thanks for pointing to these tickets. In these examples all clauses of the SpanNearQuery were SpanTermQueries, but some occured more than once. For PhraseQuery and MultiPhraseQuery and their implementation in SloppyPhraseScore this seems to be a known problem that has been solved by a special complex treatment of repetitions that I currently don't understand in detail. My current opinion: We should give up lazy iteration for the unorderd and the ordered case to solve these problems. I think it can be done and the performance peanalty should not be too big. We already iterate over all positions of all subspans. So we already have done the expensive operation of reading them. Should some more comparisons of int-values (positions) really matter so much? At least fo the ordered case I am optimistic that I could implement it efficiently. 2) Inconsistent Scoring of SpanNearQuery *) Lazy iteration means that some "redundant" matches in a document are skipped in order to have a faster matching algorithm. I am not sure how redundant was defined exactly for the idea of lazy iteration. It referred to matches with the same start posisiton somehow. As long as different matches for the first clause are concerned, they are found, but not the all matches for intermediate subclauses are regarded. Skipping matches however reduces the frequency that is computed and consequently the score. See Javadoc of phraseFreq() in SloppyPhraseScore which mention the same phenomenon. This is quite important for my use case of SpanQueries. I have different versions/variants of the same term on the same position, e.g. one with case-normalization and one without and I want a higher score if the user-query matches for more than one variant, and I use this approach for clauses of SpanNearQuery. *) In NearSpansOrdered the method width() (it is used to compute sloppy frequency in SpanScore) returns the number of gaps between the matches. If you have a perfect match it returns 0 (no sloppyness). In NearSpansUnordered it returns the length of the match, not the number of gaps. See atMatch() for the difference. The reason is probably, that (maxEndPositionCell.endPosition() - minPositionCell().startPosition() - totalSpanLength) might even become negative if matches overlap. I would prefer something like Math.max(0, (maxEndPositionCell.endPosition() - minPositionCell().startPosition() - totalSpanLength)) *) SpanOrQuery and SpanNearQuery completely ignore the scores of their subclauses (subweights are always generated as non-scoring). A SpanOrQuery should give a Score similar to a BooleanQuery, shouldn't it? As long as we have this behavior, SpanBoostQuery does not make any sense, doese it? So to my opinion the existance of SpanBoostQuery shows that others also had the idea that a nested SpanQuery should somehow use the scores of their clauses for the computation of their own score. was (Author: gol...@detego-software.de): After thoroughly reviewing the current implementations of SpanNearQuery, PhraseQuery and MultiPhraseQuery I see some problems and inconsistencies. I volunteer to fix at least some of these problems, but first I would like to have a consensus about the desired bahavior of SpanQuery. This ticket may not be the right place for such
[jira] [Comment Edited] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15466888#comment-15466888 ] Christoph Goller edited comment on LUCENE-7398 at 9/6/16 9:14 AM: -- After thoroughly reviewing the current implementations of SpanNearQuery, PhraseQuery and MultiPhraseQuery I see some problems and inconsistencies. I volunteer to fix at least some of these problems, but first I would like to have a consensus about the desired bahavior of SpanQuery. This ticket may not be the right place for such a discussion, so please point me to a better place if there is one. 1) Missing Matches caused by lazy iteration: I think lazy iteration is not a new thing in Lucene SpanNearQuery. As far as I know there never was an implementation that compared all possible combinations of subspan matches for SpanNearQuery in Lucene. So SpanNEarQuery always missed some matches. *) This ticket demonstrates missing matches for ordered SpanQuery. They are caused by subspans of SpanNearQuery having a variable match length. For these cases the lazy iteration implementation which tries to optimize the number of comparisons of subspan matches is not sufficient. *) Tim tried these examples with unorderd SpanQuery and got the same bahavior. I think this is caused by a similar kind of lazy iteration in the unordered case. *) In the unorderd case lazy iteration also causes problems if the subspans do not have variable-length matches. This is demonstrated in LUCENE-5331 and LUCENE-2861. Tim, thanks for pointing to these tickets. In these examples all clauses of the SpanNearQuery were SpanTermQueries, but some occured more than once. For PhraseQuery and MultiPhraseQuery and their implementation in SloppyPhraseScore this seems to be a known problem that has been solved by a special complex treatment of repetitions that I currently don't understand in detail. My current opinion: We should give up lazy iteration for the unorderd and the ordered case to solve these problems. I think it can be done and the performance peanalty should not be too big. We already iterate over all positions of all subspans. So we already have done the expensive operation of reading them. Should some more comparisons of int-values (positions) really matter so much? At least fo the ordered case I am optimistic that I could implement it efficiently. 2) Inconsistent Scoring of SpanNearQuery *) Lazy iteration means that some "redundant" matches in a document are skipped in order to have a faster matching algorithm. I am not sure how redundant was defined exactly for the idea of lazy iteration. It referred to matches with the same start posisiton somehow. As long as different matches for the first clause are concerned, they are found, but not the all matches for intermediate subclauses are regarded. Skipping matches however reduces the frequency that is computed and consequently the score. See Javadoc of phraseFreq() in SloppyPhraseScore which mention the same phenomenon. This is quite important for my use case of SpanQueries. I have different versions/variants of the same term on the same position, e.g. one with case-normalization and one without and I want a higher score if the user-query matches for more than one variant, and I use this approach for clauses of SpanNearQuery. *) In NearSpansOrdered the method width() (it is used to compute sloppy frequency in SpanScore) returns the number of gaps between the matches. If you have a perfect match it returns 0 (no sloppyness). In NearSpansUnordered it returns the length of the match, not the number of gaps. See atMatch() for the difference. The reason is probably, that (maxEndPositionCell.endPosition() - minPositionCell().startPosition() - totalSpanLength) might even become negative if matches overlap. I would prefer something like Math.max(0, (maxEndPositionCell.endPosition() - minPositionCell().startPosition() - totalSpanLength)) *) SpanOrQuery and SpanNearQuery completely ignore the scores of their subclauses (subweights are always generated as non-scoring). A SpanOrQuery should give a Score similar to a BooleanQuery, shouldn't it? As long as we have this behavior, SpanBoostQuery does not make any sense, doese it? So to my opinion the existance of SpanBoostQuery shows that others also had the idea that a nested SpanQuery should somehow use the scores of their clauses for the computation of their own score. was (Author: gol...@detego-software.de): After thoroughly reviewing the current implementations of SpanNearQuery, PhraseQuery and MultiPhraseQuery I see some problems and inconsistencies. I volunteer to fix at least some of these problems, but first I would like to have a consensus about the desired bahavior of SpanQuery. This ticket may not be the right place for such a discussion, so please point me to a better pl
[jira] [Commented] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15466888#comment-15466888 ] Christoph Goller commented on LUCENE-7398: -- After thoroughly reviewing the current implementations of SpanNearQuery, PhraseQuery and MultiPhraseQuery I see some problems and inconsistencies. I volunteer to fix at least some of these problems, but first I would like to have a consensus about the desired bahavior of SpanQuery. This ticket may not be the right place for such a discussion, so please point me to a better place if there is one. 1) Missing Matches caused by lazy iteration: I think lazy iteration is not a new thing in Lucene SpanNearQuery. As far as I know there never was an implementation that compared all possible combinations of subspan matches for SpanNearQuery in Lucene. So SpanNEarQuery always missed some matches. *) This ticket demonstrates missing matches for ordered SpanQuery. They are caused by subspans of SpanNearQuery having a variable match length. For these cases the lazy iteration implementation which tries to optimize the number of comparisons of subspan matches is not sufficient. *) Tim tried these examples with unorderd SpanQuery and got the same bahavior. I think this is caused by a similar kind of lazy iteration in the unordered case. *) In the unorderd case lazy iteration also causes problems if the subspans do not have variable-length matches. This is demonstrated in LUCENE-5331 and LUCENE-2861. Tim, thanks for pointing to these tickets. In these examples all clauses of the SpanNearQuery were SpanTermQueries, but some occured more than once. For PhraseQuery and MultiPhraseQuery and their implementation in SloppyPhraseScore this seems to be a known problem that has been solved by a special complex treatment of repetitions that I currently don't understand in detail. My current opinion: We should give up lazy iteration for the unorderd and the ordered case to solve these problems. I think it can be done and the performance peanalty should not be too big. We already iterate over all positions of all subspans. So we already have done the expensive operation of reading them. Should some more comparisons of int-values (positions) really matter so much? At least fo the ordered case I am optimistic that I could implement it efficiently. 2) Inconsistent Scoring of SpanNearQuery *) Lazy iteration means that some "redundant" matches in a document are skipped in order to have a faster matching algorithm. I am not sure how redundant was defined exactly for the idea of lazy iteration. It referred to matches with the same start posisiton somehow. Skpping matches however reduces the frequency that is computed and consequently the score. See Javadoc of phraseFreq() in SloppyPhraseScore which mention the same phenomenon. This is quite important for my use case of SpanQueries. I have different versions/variants of the same term on the same position, e.g. one with case-normalization and one without and I want a higher score if the user-query matches for more than one variant, and I use this approach for clauses of SpanNearQuery. *) In NearSpansOrdered the method width() (it is used to compute sloppy frequency in SpanScore) returns the number of gaps between the matches. If you have a perfect match it returns 0 (no sloppyness). In NearSpansUnordered it returns the length of the match, not the number of gaps. See atMatch() for the difference. The reason is probably, that (maxEndPositionCell.endPosition() - minPositionCell().startPosition() - totalSpanLength) might even become negative if matches overlap. I would prefer something like Math.max(0, (maxEndPositionCell.endPosition() - minPositionCell().startPosition() - totalSpanLength)) *) SpanOrQuery and SpanNearQuery completely ignore the scores of their subclauses (subweights are always generated as non-scoring). A SpanOrQuery should give a Score similar to a BooleanQuery, shouldn't it? As long as we have this behavior, SpanBoostQuery does not make any sense, doese it? So to my opinion the existance of SpanBoostQuery shows that others also had the idea that a nested SpanQuery should somehow use the scores of their clauses for the computation of their own score. > Nested Span Queries are buggy > - > > Key: LUCENE-7398 > URL: https://issues.apache.org/jira/browse/LUCENE-7398 > Project: Lucene - Core > Issue Type: Bug > Components: core/search >Affects Versions: 5.5, 6.x >Reporter: Christoph Goller >Assignee: Alan Woodward >Priority: Critical > Attachments: LUCENE-7398-20160814.patch, LUCENE-7398.patch, > LUCENE-7398.patch, TestSpanCollection.java > > > Example for a nested SpanQuery that is not working: > Document: Human Genome Organiza
[jira] [Commented] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15465336#comment-15465336 ] Christoph Goller commented on LUCENE-7398: -- Good idea to try the nested tests from TestSpanCollection for the unordered case. The example from LUCENE-5331 shows the problems of incomplete backtracking (not comparing all combinations of span matches of all subspans) for the unordered case. In the ordered case we only have a problem with spans that have matches of different lenght, in the unorderd case we also see a problem with overlapping span-matches, even if they all have length 1. > Nested Span Queries are buggy > - > > Key: LUCENE-7398 > URL: https://issues.apache.org/jira/browse/LUCENE-7398 > Project: Lucene - Core > Issue Type: Bug > Components: core/search >Affects Versions: 5.5, 6.x >Reporter: Christoph Goller >Assignee: Alan Woodward >Priority: Critical > Attachments: LUCENE-7398-20160814.patch, LUCENE-7398.patch, > LUCENE-7398.patch, TestSpanCollection.java > > > Example for a nested SpanQuery that is not working: > Document: Human Genome Organization , HUGO , is trying to coordinate gene > mapping research worldwide. > Query: spanNear([body:coordinate, spanOr([spanNear([body:gene, body:mapping], > 0, true), body:gene]), body:research], 0, true) > The query should match "coordinate gene mapping research" as well as > "coordinate gene research". It does not match "coordinate gene mapping > research" with Lucene 5.5 or 6.1, it did however match with Lucene 4.10.4. It > probably stopped working with the changes on SpanQueries in 5.3. I will > attach a unit test that shows the problem. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Comment Edited] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15465301#comment-15465301 ] Christoph Goller edited comment on LUCENE-7398 at 9/5/16 4:07 PM: -- Paul's 20160814 patch almost convinced me. Unfortunately, it does not fix the case when an intermediate span has a longer match that reduces overall sloppyness but overlaps with a match of a subsequent span and consequently requires advancing the subsequent span. Here is an example Document: w1 w2 w3 w4 w5 near/0(w1, or(w2, near/0(w2, w3, w4)), or(w5, near/0(w4, w5))) Add the following code to the end of TestSpanCollection.testNestedNearQuery() {code} SpanNearQuery q234 = new SpanNearQuery(new SpanQuery[]{q2, q3, q4}, 0, true); SpanOrQuery q2234 = new SpanOrQuery(q2, q234); SpanTermQuery p5 = new SpanTermQuery(new Term(FIELD, "w5")); SpanNearQuery q45 = new SpanNearQuery(new SpanQuery[]{q4, p5}, 0, true); SpanOrQuery q455 = new SpanOrQuery(q45, p5); SpanNearQuery q1q2234q445 = new SpanNearQuery(new SpanQuery[]{q1, q2234, q455}, 0, true); spans = q1q2234q445.createWeight(searcher, false, 1f).getSpans(searcher.getIndexReader().leaves().get(0),SpanWeight.Postings.POSITIONS); assertEquals(0, spans.advance(0)); {code} I think we can only fix it if we get give up lazy iteration. I don't think this is so bad for performance. If we implement a clever caching for positions in spans a complete backtracking would only consist of making a few additional int-comparisons. The expensive operation is iterating over all span positions (IO) and we do this already in advancePosition(Spans, int), aren't we. was (Author: gol...@detego-software.de): Paul's fix almost convinced me. Unfortunately, it does not fix the case when an intermediate span has a longer match that reduces overall sloppyness but overlaps with a match of a subsequent span and consequently requires advancing the subsequent span. Here is an example Document: w1 w2 w3 w4 w5 near/0(w1, or(w2, near/0(w2, w3, w4)), or(w5, near/0(w4, w5))) Add the following code to the end of TestSpanCollection.testNestedNearQuery() {code} SpanNearQuery q234 = new SpanNearQuery(new SpanQuery[]{q2, q3, q4}, 0, true); SpanOrQuery q2234 = new SpanOrQuery(q2, q234); SpanTermQuery p5 = new SpanTermQuery(new Term(FIELD, "w5")); SpanNearQuery q45 = new SpanNearQuery(new SpanQuery[]{q4, p5}, 0, true); SpanOrQuery q455 = new SpanOrQuery(q45, p5); SpanNearQuery q1q2234q445 = new SpanNearQuery(new SpanQuery[]{q1, q2234, q455}, 0, true); spans = q1q2234q445.createWeight(searcher, false, 1f).getSpans(searcher.getIndexReader().leaves().get(0),SpanWeight.Postings.POSITIONS); assertEquals(0, spans.advance(0)); {code} I think we can only fix it if we get give up lazy iteration. I don't think this is so bad for performance. If we implement a clever caching for positions in spans a complete backtracking would only consist of making a few additional int-comparisons. The expensive operation is iterating over all span positions (IO) and we do this already in advancePosition(Spans, int), aren't we. > Nested Span Queries are buggy > - > > Key: LUCENE-7398 > URL: https://issues.apache.org/jira/browse/LUCENE-7398 > Project: Lucene - Core > Issue Type: Bug > Components: core/search >Affects Versions: 5.5, 6.x >Reporter: Christoph Goller >Assignee: Alan Woodward >Priority: Critical > Attachments: LUCENE-7398-20160814.patch, LUCENE-7398.patch, > LUCENE-7398.patch, TestSpanCollection.java > > > Example for a nested SpanQuery that is not working: > Document: Human Genome Organization , HUGO , is trying to coordinate gene > mapping research worldwide. > Query: spanNear([body:coordinate, spanOr([spanNear([body:gene, body:mapping], > 0, true), body:gene]), body:research], 0, true) > The query should match "coordinate gene mapping research" as well as > "coordinate gene research". It does not match "coordinate gene mapping > research" with Lucene 5.5 or 6.1, it did however match with Lucene 4.10.4. It > probably stopped working with the changes on SpanQueries in 5.3. I will > attach a unit test that shows the problem. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15465301#comment-15465301 ] Christoph Goller commented on LUCENE-7398: -- Paul's fix almost convinced me. Unfortunately, it does not fix the case when an intermediate span has a longer match that reduces overall sloppyness but overlaps with a match of a subsequent span and consequently requires advancing the subsequent span. Here is an example Document: w1 w2 w3 w4 w5 near/0(w1, or(w2, near/0(w2, w3, w4)), or(w5, near/0(w4, w5))) Add the following code to the end of TestSpanCollection.testNestedNearQuery() {code} SpanNearQuery q234 = new SpanNearQuery(new SpanQuery[]{q2, q3, q4}, 0, true); SpanOrQuery q2234 = new SpanOrQuery(q2, q234); SpanTermQuery p5 = new SpanTermQuery(new Term(FIELD, "w5")); SpanNearQuery q45 = new SpanNearQuery(new SpanQuery[]{q4, p5}, 0, true); SpanOrQuery q455 = new SpanOrQuery(q45, p5); SpanNearQuery q1q2234q445 = new SpanNearQuery(new SpanQuery[]{q1, q2234, q455}, 0, true); spans = q1q2234q445.createWeight(searcher, false, 1f).getSpans(searcher.getIndexReader().leaves().get(0),SpanWeight.Postings.POSITIONS); assertEquals(0, spans.advance(0)); {code} I think we can only fix it if we get give up lazy iteration. I don't think this is so bad for performance. If we implement a clever caching for positions in spans a complete backtracking would only consist of making a few additional int-comparisons. The expensive operation is iterating over all span positions (IO) and we do this already in advancePosition(Spans, int), aren't we. > Nested Span Queries are buggy > - > > Key: LUCENE-7398 > URL: https://issues.apache.org/jira/browse/LUCENE-7398 > Project: Lucene - Core > Issue Type: Bug > Components: core/search >Affects Versions: 5.5, 6.x >Reporter: Christoph Goller >Assignee: Alan Woodward >Priority: Critical > Attachments: LUCENE-7398-20160814.patch, LUCENE-7398.patch, > LUCENE-7398.patch, TestSpanCollection.java > > > Example for a nested SpanQuery that is not working: > Document: Human Genome Organization , HUGO , is trying to coordinate gene > mapping research worldwide. > Query: spanNear([body:coordinate, spanOr([spanNear([body:gene, body:mapping], > 0, true), body:gene]), body:research], 0, true) > The query should match "coordinate gene mapping research" as well as > "coordinate gene research". It does not match "coordinate gene mapping > research" with Lucene 5.5 or 6.1, it did however match with Lucene 4.10.4. It > probably stopped working with the changes on SpanQueries in 5.3. I will > attach a unit test that shows the problem. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (LUCENE-5331) nested SpanNearQuery with repeating groups does not find match
[ https://issues.apache.org/jira/browse/LUCENE-5331?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15464619#comment-15464619 ] Christoph Goller commented on LUCENE-5331: -- As LUCENE-2861 the problem is caused by overlappiung matches for d, b, and c and an incomplete backtracking mechanism in unordered SpanQuery. > nested SpanNearQuery with repeating groups does not find match > -- > > Key: LUCENE-5331 > URL: https://issues.apache.org/jira/browse/LUCENE-5331 > Project: Lucene - Core > Issue Type: Bug >Reporter: Jerry Zhou > Attachments: NestedSpanNearTest.java, > NestedSpanNearTest_20160902.patch > > > Nested spanNear queries do not work in some cases when repeating groups are > in the query. > Test case is attached ... -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (LUCENE-2861) Search doesn't return document via query
[ https://issues.apache.org/jira/browse/LUCENE-2861?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15464613#comment-15464613 ] Christoph Goller commented on LUCENE-2861: -- The problem is caused by overlapping matches within spanNear2. The first match for spanNear2 matches "intended message" and the second "message" to the same "message" in the text so that the match for "addressed" ist to far away. One possible fix would forbid overlapping matches or add aspecial very compley treatment like in SloppyPhraseScore. I think it would be better to give up lazy backtracking and implement a correct backtracking (see LUCENE-7398). > Search doesn't return document via query > > > Key: LUCENE-2861 > URL: https://issues.apache.org/jira/browse/LUCENE-2861 > Project: Lucene - Core > Issue Type: Bug > Components: core/search >Affects Versions: 2.9.1, 2.9.4, 3.0.3 > Environment: Doesn't depend on enviroment >Reporter: Zenoviy Veres > > The query doesn't return document that contain all words from query in > correct order. > The issue might be within mechanism how do SpanQuerys actually match results > (http://www.lucidimagination.com/blog/2009/07/18/the-spanquery/) > Please refer for details below. The example text wasn't passed through > snowball analyzer, however the issue exists after analyzing too > Query: > (intend within 3 of message) within 5 of message within 3 of addressed. > Text within document: > The contents of this e-mail message and > any attachments are intended solely for the > addressee(s) and may contain confidential > and/or legally privileged information. If you > are not the intended recipient of this message > or if this message has been addressed to you > in error, please immediately alert the sender > by reply e-mail and then delete this message > and any attachments > Result query: > SpanNearQuery spanNear = new SpanNearQuery(new SpanQuery[] { > new SpanTermQuery(new Term(BODY, "intended")), > new SpanTermQuery(new Term(BODY, "message"))}, > 4, > false); > SpanNearQuery spanNear2 = new SpanNearQuery(new SpanQuery[] > {spanNear, new SpanTermQuery(new Term(BODY, "message"))}, 5, false); > SpanNearQuery spanNear3 = new SpanNearQuery(new SpanQuery[] > {spanNear2, new SpanTermQuery(new Term(BODY, "addressed"))}, 3, false); -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Commented] (LUCENE-5396) SpanNearQuery returns single term spans
[ https://issues.apache.org/jira/browse/LUCENE-5396?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15464535#comment-15464535 ] Christoph Goller commented on LUCENE-5396: -- Is this a bug or desired bahavior? For me it is at least an acceptable behavior. I like the behavior of unordered SpanNearQuery to match if clauses overlap or match at the same position. and it would be quite difficult to find out if two clauses match at the same index term or only at the same position. background: I am using a component for word decomposition. This might be a very rare case for English but it is a much more common phenomen for German and Dutch. The two compound parts of "wallpaper" (wall and paper) go into the same index position as wallpaper. I am using spanNear([wall, paper], 0, false) to search for wallpaper and expect matches for "wallpaper" as well as for "wall paper". So far we do not have a proper definition of what SpanQueries should do and the only way to find out what they currently do is to look into the code. I think the current behavior is not very consistent. I will present some of my insights and ideas in LUCENE-7398. > SpanNearQuery returns single term spans > --- > > Key: LUCENE-5396 > URL: https://issues.apache.org/jira/browse/LUCENE-5396 > Project: Lucene - Core > Issue Type: Bug > Components: core/search >Reporter: Piotr Pęzik > > Let's assume we have an index with two documents: > 1. contents: "test bunga bunga test" > 2. contents: "test bunga test" > We run two SpanNearQueries against this index: > 1. spanNear([contents:bunga, contents:bunga], 0, true) > 2. spanNear([contents:bunga, contents:bunga], 0, false) > For the first query we get 1 hit. The first document in the example above > gets matched and the second one doesn't. This make sense, because we want the > term "bunga" followed by another "bunga" here. > However, both documents get matched by the second query. This is also > problematic in cases where we have duplicate terms in longer (unordered) > spannear queries, e. g.: unordered 'A B A' will match spans such as 'A B' or > 'B A'. > A complete example follows. > - > import org.apache.lucene.analysis.Analyzer; > import org.apache.lucene.analysis.standard.StandardAnalyzer; > import org.apache.lucene.document.Document; > import org.apache.lucene.document.TextField; > import org.apache.lucene.index.DirectoryReader; > import org.apache.lucene.index.IndexWriter; > import org.apache.lucene.index.IndexWriterConfig; > import org.apache.lucene.index.Term; > import org.apache.lucene.search.IndexSearcher; > import org.apache.lucene.search.TopDocs; > import org.apache.lucene.search.spans.SpanNearQuery; > import org.apache.lucene.search.spans.SpanQuery; > import org.apache.lucene.search.spans.SpanTermQuery; > import org.apache.lucene.store.Directory; > import org.apache.lucene.store.FSDirectory; > import org.apache.lucene.store.RAMDirectory; > import org.apache.lucene.util.Version; > import java.io.StringReader; > import static org.junit.Assert.assertEquals; > class SpansBug { > public static void main(String [] args) throws Exception { > Directory dir = new RAMDirectory(); > Analyzer analyzer = new StandardAnalyzer(Version.LUCENE_45); > IndexWriterConfig iwc = new IndexWriterConfig(Version.LUCENE_45, > analyzer); > IndexWriter writer = new IndexWriter(dir, iwc); > String contents = "contents"; > Document doc1 = new Document(); > doc1.add(new TextField(contents, new StringReader("test bunga bunga > test"))); > Document doc2 = new Document(); > doc2.add(new TextField(contents, new StringReader("test bunga > test"))); > writer.addDocument(doc1); > writer.addDocument(doc2); > writer.commit(); > IndexSearcher searcher = new IndexSearcher(DirectoryReader.open(dir)); > SpanQuery stq1 = new SpanTermQuery(new Term(contents,"bunga")); > SpanQuery stq2 = new SpanTermQuery(new Term(contents,"bunga")); > SpanQuery [] spqa = new SpanQuery[]{stq1,stq2}; > SpanNearQuery spanQ1 = new SpanNearQuery(spqa,0, true); > SpanNearQuery spanQ2 = new SpanNearQuery(spqa,0, false); > System.out.println(spanQ1); > TopDocs tdocs1 = searcher.search(spanQ1,10); > assertEquals(tdocs1.totalHits ,1); > System.out.println(spanQ2); > TopDocs tdocs2 = searcher.search(spanQ2,10); > //I'd expect one hit here: > assertEquals(tdocs2.totalHits ,1); // Assertion fails > } > } -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.
[jira] [Commented] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15406059#comment-15406059 ] Christoph Goller commented on LUCENE-7398: -- The whole idea of the patch is to change the order of the matches returned by SpanOrQuery. {code} SpanTermQuery q2 = new SpanTermQuery(new Term(FIELD, "w2")); SpanTermQuery q3 = new SpanTermQuery(new Term(FIELD, "w3")); SpanNearQuery q23 = new SpanNearQuery(new SpanQuery[]{q2, q3}, 0, true); SpanOrQuery q223 = new SpanOrQuery(q2, q23); {code} For a document containing "w1 w2 w3 w4" query q223 now returns as first match "w2 w3" (the longer one) and then "w2" while formerly it was the other way round. Both matches have the same start position, but different end positions and the contract about spans says that if start positions equal we first get the match with the lower end position (Javadoc of spans). > Nested Span Queries are buggy > - > > Key: LUCENE-7398 > URL: https://issues.apache.org/jira/browse/LUCENE-7398 > Project: Lucene - Core > Issue Type: Bug > Components: core/search >Affects Versions: 5.5, 6.x >Reporter: Christoph Goller >Assignee: Alan Woodward >Priority: Critical > Attachments: LUCENE-7398.patch, LUCENE-7398.patch, > TestSpanCollection.java > > > Example for a nested SpanQuery that is not working: > Document: Human Genome Organization , HUGO , is trying to coordinate gene > mapping research worldwide. > Query: spanNear([body:coordinate, spanOr([spanNear([body:gene, body:mapping], > 0, true), body:gene]), body:research], 0, true) > The query should match "coordinate gene mapping research" as well as > "coordinate gene research". It does not match "coordinate gene mapping > research" with Lucene 5.5 or 6.1, it did however match with Lucene 4.10.4. It > probably stopped working with the changes on SpanQueries in 5.3. I will > attach a unit test that shows the problem. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Comment Edited] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15405903#comment-15405903 ] Christoph Goller edited comment on LUCENE-7398 at 8/3/16 1:21 PM: -- After thoroughly looking into SpanQueries my conclusion is, that we have a fundamental problem in the implementation of SpanNearQuery. The problem is not new, it probably existed already in the first version of SpanQueries which as far as I know were implemented by Doug Cutting himself. I remember some attempts to describe in which cases SpanQueries work correctly and in which they do not (discussions about overlapping), but those explanations and definitions were never completely convincing for me. My best guess: NearSpansOrdered and NearSpansUnordered currently are only correct if for each clause of the SpanQuery we can guarantee, that all its matches have the same length. In this case it is clear that (for the ordered case) if a match is too long (sloppy) we can skip to the first clause and call nextPosition. No alternative matches of intermediate clauses could improve the overall match. If we have clauses with varying match length (SpanOr or SpanNear with sloppyness) we would have to backtrack to intermediate clauses and check whether there are e.g. longer matches that could reduce the overall match length. Pauls last test case shows that even a match of the second clause that advances its position can reduce the overall lenght if it is longer himself. A match of an intermediate clause at an advanced position could be considerably shorter than its first match requiring a reset of the spans of following clauses. To my opinion this bug can only be fixed by implementing a backtracking search on the subspans that also requires a limited possibilitxy to reposition Spans to previous positions. By the way, shrinkToAfterShortestMatch() in NearSpansOrdered of Lucene 4_10_4 provided a kind of backtracking which was the reason why my queries worked in elasticsearch 1.7.x. However, I think the implementation also did not solve all cases: {code} /** The subSpans are ordered in the same doc, so there is a possible match. * Compute the slop while making the match as short as possible by advancing * all subSpans except the last one in reverse order. */ private boolean shrinkToAfterShortestMatch() throws IOException { matchStart = subSpans[subSpans.length - 1].start(); matchEnd = subSpans[subSpans.length - 1].end(); Set possibleMatchPayloads = new HashSet<>(); if (subSpans[subSpans.length - 1].isPayloadAvailable()) { possibleMatchPayloads.addAll(subSpans[subSpans.length - 1].getPayload()); } Collection possiblePayload = null; int matchSlop = 0; int lastStart = matchStart; int lastEnd = matchEnd; for (int i = subSpans.length - 2; i >= 0; i--) { Spans prevSpans = subSpans[i]; if (collectPayloads && prevSpans.isPayloadAvailable()) { Collection payload = prevSpans.getPayload(); possiblePayload = new ArrayList<>(payload.size()); possiblePayload.addAll(payload); } int prevStart = prevSpans.start(); int prevEnd = prevSpans.end(); while (true) { // Advance prevSpans until after (lastStart, lastEnd) if (! prevSpans.next()) { inSameDoc = false; more = false; break; // Check remaining subSpans for final match. } else if (matchDoc != prevSpans.doc()) { inSameDoc = false; // The last subSpans is not advanced here. break; // Check remaining subSpans for last match in this document. } else { int ppStart = prevSpans.start(); int ppEnd = prevSpans.end(); // Cannot avoid invoking .end() if (! docSpansOrderedNonOverlap(ppStart, ppEnd, lastStart, lastEnd)) { break; // Check remaining subSpans. } else { // prevSpans still before (lastStart, lastEnd) prevStart = ppStart; prevEnd = ppEnd; if (collectPayloads && prevSpans.isPayloadAvailable()) { Collection payload = prevSpans.getPayload(); possiblePayload = new ArrayList<>(payload.size()); possiblePayload.addAll(payload); } } } } if (collectPayloads && possiblePayload != null) { possibleMatchPayloads.addAll(possiblePayload); } assert prevStart <= matchStart; if (matchStart > prevEnd) { // Only non overlapping spans add to slop. matchSlop += (matchStart - prevEnd); } /* Do not break on (matchSlop > allowedSlop) here to make sure * that subSpans[0] is advanced after the match, if any. */ matchStart = prevStart; lastStart = prevStart; lastEnd = prevEnd; } boolean match = matchSlop <= allowedSlop;
[jira] [Comment Edited] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15405903#comment-15405903 ] Christoph Goller edited comment on LUCENE-7398 at 8/3/16 1:20 PM: -- After thoroughly looking into SpanQueries my conclusion is, that we have a fundamental problem in the implementation of SpanNearQuery. The problem is not new, it probably existed already in the first version of SpanQueries which as far as I know were implemented by Doug Cutting himself. I remember some attempts to describe in which cases SpanQueries work correctly and in which they do not (discussions about overlapping), but those explanations and definitions were never completely convincing for me. My best guess: NearSpansOrdered and NearSpansUnordered currently are only correct if for each clause of the SpanQuery we can guarantee, that all its matches have the same length. In this case it is clear that (for the ordered case) if a match is too long (sloppy) we can skip to the first clause and call nextPosition. No alternative matches of intermediate clauses could improve the overall match. If we have clauses with varying match length (SpanOr or SpanNear with sloppyness) we would have to backtrack to intermediate clauses and check whether there are e.g. longer matches that could reduce the overall match length. Pauls last test case shows that even a match of the second clause that advances its position can reduce the overall lenght if it is longer himself. A match of an intermediate clause at an advanced position could be considerably shorter than its first match requiring a reset of the spans of following clauses. To my opinion this bug can only be fixed by implementing a backtracking search on the subspans that also requires a limited possibilitxy to reposition Spans to previous positions. By the way, shrinkToAfterShortestMatch() in NearSpansOrdered of Lucene 4_10_4 provided a kind of backtracking which was the reason why my queries worked in elasticsearch 1.7.x. However, I think the implementation also did not solve all cases: {code} /** The subSpans are ordered in the same doc, so there is a possible match. * Compute the slop while making the match as short as possible by advancing * all subSpans except the last one in reverse order. */ private boolean shrinkToAfterShortestMatch() throws IOException { matchStart = subSpans[subSpans.length - 1].start(); matchEnd = subSpans[subSpans.length - 1].end(); Set possibleMatchPayloads = new HashSet<>(); if (subSpans[subSpans.length - 1].isPayloadAvailable()) { possibleMatchPayloads.addAll(subSpans[subSpans.length - 1].getPayload()); } Collection possiblePayload = null; int matchSlop = 0; int lastStart = matchStart; int lastEnd = matchEnd; for (int i = subSpans.length - 2; i >= 0; i--) { Spans prevSpans = subSpans[i]; if (collectPayloads && prevSpans.isPayloadAvailable()) { Collection payload = prevSpans.getPayload(); possiblePayload = new ArrayList<>(payload.size()); possiblePayload.addAll(payload); } int prevStart = prevSpans.start(); int prevEnd = prevSpans.end(); while (true) { // Advance prevSpans until after (lastStart, lastEnd) if (! prevSpans.next()) { inSameDoc = false; more = false; break; // Check remaining subSpans for final match. } else if (matchDoc != prevSpans.doc()) { inSameDoc = false; // The last subSpans is not advanced here. break; // Check remaining subSpans for last match in this document. } else { int ppStart = prevSpans.start(); int ppEnd = prevSpans.end(); // Cannot avoid invoking .end() if (! docSpansOrderedNonOverlap(ppStart, ppEnd, lastStart, lastEnd)) { break; // Check remaining subSpans. } else { // prevSpans still before (lastStart, lastEnd) prevStart = ppStart; prevEnd = ppEnd; if (collectPayloads && prevSpans.isPayloadAvailable()) { Collection payload = prevSpans.getPayload(); possiblePayload = new ArrayList<>(payload.size()); possiblePayload.addAll(payload); } } } } if (collectPayloads && possiblePayload != null) { possibleMatchPayloads.addAll(possiblePayload); } assert prevStart <= matchStart; if (matchStart > prevEnd) { // Only non overlapping spans add to slop. matchSlop += (matchStart - prevEnd); } /* Do not break on (matchSlop > allowedSlop) here to make sure * that subSpans[0] is advanced after the match, if any. */ matchStart = prevStart; lastStart = prevStart; lastEnd = prevEnd; } boolean match = matchSlop <= allowedSlop;
[jira] [Comment Edited] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15405903#comment-15405903 ] Christoph Goller edited comment on LUCENE-7398 at 8/3/16 1:19 PM: -- After thoroughly looking into SpanQueries my conclusion is, that we have a fundamental problem in the implementation of SpanNearQuery. The problem is not new, it probably existed already in the first version of SpanQueries which as far as I know were implemented by Doug Cutting himself. I remember some attempts to describe in which cases SpanQueries work correctly and in which they do not (discussions about overlapping), but those explanations and definitions were never completely convincing for me. My best guess: NearSpansOrdered and NearSpansUnordered currently are only correct if for each clause of the SpanQuery we can guarantee, that all its matches have the same length. In this case it is clear that (for the ordered case) if a match is too long (sloppy) we can skip to the first clause and call nextPosition. No alternative matches of intermediate clauses could improve the overall match. If we have clauses with varying match length (SpanOr or SpanNear with sloppyness) we would have to backtrack to intermediate clauses and check whether there are e.g. longer matches that could reduce the overall match length. Pauls last test case shows that even a match of the second clause that advances its position can reduce the overall lenght if it is longer himself. A match of an intermediate clause at an advanced position could be considerably shorter than its first match requiring a reset of the spans of following clauses. To my opinion this bug can only be fixed by implementing a backtracking search on the subspans that also requires a limited possibilitxy to reposion Spans to previous positions. By the way, shrinkToAfterShortestMatch() in NearSpansOrdered of Lucene 4_10_4 provided a kind of backtracking which was the reason why my queries worked in elasticsearch 1.7.x. However, I think the implementation also did not solve all cases: {code} /** The subSpans are ordered in the same doc, so there is a possible match. * Compute the slop while making the match as short as possible by advancing * all subSpans except the last one in reverse order. */ private boolean shrinkToAfterShortestMatch() throws IOException { matchStart = subSpans[subSpans.length - 1].start(); matchEnd = subSpans[subSpans.length - 1].end(); Set possibleMatchPayloads = new HashSet<>(); if (subSpans[subSpans.length - 1].isPayloadAvailable()) { possibleMatchPayloads.addAll(subSpans[subSpans.length - 1].getPayload()); } Collection possiblePayload = null; int matchSlop = 0; int lastStart = matchStart; int lastEnd = matchEnd; for (int i = subSpans.length - 2; i >= 0; i--) { Spans prevSpans = subSpans[i]; if (collectPayloads && prevSpans.isPayloadAvailable()) { Collection payload = prevSpans.getPayload(); possiblePayload = new ArrayList<>(payload.size()); possiblePayload.addAll(payload); } int prevStart = prevSpans.start(); int prevEnd = prevSpans.end(); while (true) { // Advance prevSpans until after (lastStart, lastEnd) if (! prevSpans.next()) { inSameDoc = false; more = false; break; // Check remaining subSpans for final match. } else if (matchDoc != prevSpans.doc()) { inSameDoc = false; // The last subSpans is not advanced here. break; // Check remaining subSpans for last match in this document. } else { int ppStart = prevSpans.start(); int ppEnd = prevSpans.end(); // Cannot avoid invoking .end() if (! docSpansOrderedNonOverlap(ppStart, ppEnd, lastStart, lastEnd)) { break; // Check remaining subSpans. } else { // prevSpans still before (lastStart, lastEnd) prevStart = ppStart; prevEnd = ppEnd; if (collectPayloads && prevSpans.isPayloadAvailable()) { Collection payload = prevSpans.getPayload(); possiblePayload = new ArrayList<>(payload.size()); possiblePayload.addAll(payload); } } } } if (collectPayloads && possiblePayload != null) { possibleMatchPayloads.addAll(possiblePayload); } assert prevStart <= matchStart; if (matchStart > prevEnd) { // Only non overlapping spans add to slop. matchSlop += (matchStart - prevEnd); } /* Do not break on (matchSlop > allowedSlop) here to make sure * that subSpans[0] is advanced after the match, if any. */ matchStart = prevStart; lastStart = prevStart; lastEnd = prevEnd; } boolean match = matchSlop <= allowedSlop;
[jira] [Comment Edited] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15405903#comment-15405903 ] Christoph Goller edited comment on LUCENE-7398 at 8/3/16 1:18 PM: -- After thoroughly looking into SpanQueries my conclusion is, that we have a fundamental problem in the implementation of SpanNearQuery. The problem is not new, it probably existed already in the first version of SpanQueries which as far as I know were implemented by Doug Cutting himself. I remember some attempts to describe in which cases SpanQueries work correctly and in which they do not (discussions about overlapping), but those explanations and definitions were never completely convincing for me. My best guess: NearSpansOrdered and NearSpansUnordered currently are only correct if for each clause of the SpanQuery we can guarantee, that all its matches have the same length. In this case it is clear that (for the ordered case) if a match is too long (sloppy) we can skip to the first clause and call nextPosition. No alternative matches of intermediate clauses could improve the overall match. If we have clauses with varying match length (SpanOr or SpanNear with sloppyness) we would have to backtrack to intermediate clauses and check whether there are e.g. longer matches that could reduce the overall match length. Pauls last test case shows that even a match of the second clause that advances its position can reduce the overall lenght if it is longer himnself. A match of an intermediate clause at an advanced position could be considerably shorter than its first match requiring a reset of the spans of following clauses. To my opinion this bug can only be fixed by implementing a backtracking search on the subspans that also requires a limited possibilitxy to reposion Spans to previous positions. By the way, shrinkToAfterShortestMatch() in NearSpansOrdered of Lucene 4_10_4 provided a kind of backtracking which was the reason why my queries worked in elasticsearch 1.7.x. However, I think the implementation also did not solve all cases: {code} /** The subSpans are ordered in the same doc, so there is a possible match. * Compute the slop while making the match as short as possible by advancing * all subSpans except the last one in reverse order. */ private boolean shrinkToAfterShortestMatch() throws IOException { matchStart = subSpans[subSpans.length - 1].start(); matchEnd = subSpans[subSpans.length - 1].end(); Set possibleMatchPayloads = new HashSet<>(); if (subSpans[subSpans.length - 1].isPayloadAvailable()) { possibleMatchPayloads.addAll(subSpans[subSpans.length - 1].getPayload()); } Collection possiblePayload = null; int matchSlop = 0; int lastStart = matchStart; int lastEnd = matchEnd; for (int i = subSpans.length - 2; i >= 0; i--) { Spans prevSpans = subSpans[i]; if (collectPayloads && prevSpans.isPayloadAvailable()) { Collection payload = prevSpans.getPayload(); possiblePayload = new ArrayList<>(payload.size()); possiblePayload.addAll(payload); } int prevStart = prevSpans.start(); int prevEnd = prevSpans.end(); while (true) { // Advance prevSpans until after (lastStart, lastEnd) if (! prevSpans.next()) { inSameDoc = false; more = false; break; // Check remaining subSpans for final match. } else if (matchDoc != prevSpans.doc()) { inSameDoc = false; // The last subSpans is not advanced here. break; // Check remaining subSpans for last match in this document. } else { int ppStart = prevSpans.start(); int ppEnd = prevSpans.end(); // Cannot avoid invoking .end() if (! docSpansOrderedNonOverlap(ppStart, ppEnd, lastStart, lastEnd)) { break; // Check remaining subSpans. } else { // prevSpans still before (lastStart, lastEnd) prevStart = ppStart; prevEnd = ppEnd; if (collectPayloads && prevSpans.isPayloadAvailable()) { Collection payload = prevSpans.getPayload(); possiblePayload = new ArrayList<>(payload.size()); possiblePayload.addAll(payload); } } } } if (collectPayloads && possiblePayload != null) { possibleMatchPayloads.addAll(possiblePayload); } assert prevStart <= matchStart; if (matchStart > prevEnd) { // Only non overlapping spans add to slop. matchSlop += (matchStart - prevEnd); } /* Do not break on (matchSlop > allowedSlop) here to make sure * that subSpans[0] is advanced after the match, if any. */ matchStart = prevStart; lastStart = prevStart; lastEnd = prevEnd; } boolean match = matchSlop <= allowedSlop;
[jira] [Commented] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15405903#comment-15405903 ] Christoph Goller commented on LUCENE-7398: -- After thoroughly looking into SpanQueries my conclusion is, that we have a fundamental problem in the implementation of SpanNearQuery. The problem is not new, it probably existed already in the first version of SpanQueries which as far as I know were implemented by Doug Cutting himself. I remember some attempts to describe in which cases SpanQueries work correctly and in which they do not (discussions about overlapping), but those explanations and definitions were never completely convincing for me. My best guess: NearSpansOrdered and NearSpansUnordered currently are only correct if for each clause of the SpanQuery we can guarantee, that all its matches have the same length. In this case it is clear that (for the ordered case) if a match is too long (sloppy) we can skip to the first clause and call nextPosition. No alternative matches of intermediate clauses could improve the overall match. It we have clauses with varying match length (SpanOr or SpanNear with sloppyness) we would have to backtrack to intermediate clauses and check whether there are e.g. longer matches that could reduce the overall match length. Pauls last test case shows that even a match of the second clause that advances its position can reduce the overall lenght if it is longer himnself. A match of an intermediate clause at an advanced position could be considerably shorter than its first match requiring a reset of the spans of following clauses. To my opinion this bug can only be fixed by implementing a backtracking search on the subspans that also requires a limited possibilitxy to reposion Spans to previous positions. By the way, shrinkToAfterShortestMatch() in NearSpansOrdered of Lucene 4_10_4 provided a kind of backtracking which was the reason why my queries worked in elasticsearch 1.7.x. However, I think the implementation also did not solve all cases: {code} /** The subSpans are ordered in the same doc, so there is a possible match. * Compute the slop while making the match as short as possible by advancing * all subSpans except the last one in reverse order. */ private boolean shrinkToAfterShortestMatch() throws IOException { matchStart = subSpans[subSpans.length - 1].start(); matchEnd = subSpans[subSpans.length - 1].end(); Set possibleMatchPayloads = new HashSet<>(); if (subSpans[subSpans.length - 1].isPayloadAvailable()) { possibleMatchPayloads.addAll(subSpans[subSpans.length - 1].getPayload()); } Collection possiblePayload = null; int matchSlop = 0; int lastStart = matchStart; int lastEnd = matchEnd; for (int i = subSpans.length - 2; i >= 0; i--) { Spans prevSpans = subSpans[i]; if (collectPayloads && prevSpans.isPayloadAvailable()) { Collection payload = prevSpans.getPayload(); possiblePayload = new ArrayList<>(payload.size()); possiblePayload.addAll(payload); } int prevStart = prevSpans.start(); int prevEnd = prevSpans.end(); while (true) { // Advance prevSpans until after (lastStart, lastEnd) if (! prevSpans.next()) { inSameDoc = false; more = false; break; // Check remaining subSpans for final match. } else if (matchDoc != prevSpans.doc()) { inSameDoc = false; // The last subSpans is not advanced here. break; // Check remaining subSpans for last match in this document. } else { int ppStart = prevSpans.start(); int ppEnd = prevSpans.end(); // Cannot avoid invoking .end() if (! docSpansOrderedNonOverlap(ppStart, ppEnd, lastStart, lastEnd)) { break; // Check remaining subSpans. } else { // prevSpans still before (lastStart, lastEnd) prevStart = ppStart; prevEnd = ppEnd; if (collectPayloads && prevSpans.isPayloadAvailable()) { Collection payload = prevSpans.getPayload(); possiblePayload = new ArrayList<>(payload.size()); possiblePayload.addAll(payload); } } } } if (collectPayloads && possiblePayload != null) { possibleMatchPayloads.addAll(possiblePayload); } assert prevStart <= matchStart; if (matchStart > prevEnd) { // Only non overlapping spans add to slop. matchSlop += (matchStart - prevEnd); } /* Do not break on (matchSlop > allowedSlop) here to make sure * that subSpans[0] is advanced after the match, if any. */ matchStart = prevStart; lastStart = prevStart; lastEnd = prevEnd; } boolean match = matchSlop <= allowedSlop; if(collectPayloads && match && possibl
[jira] [Comment Edited] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15399414#comment-15399414 ] Christoph Goller edited comment on LUCENE-7398 at 7/29/16 2:40 PM: --- Please find attatched an extended TestSpanCollection.java for Lucene 6.1 that shows the problem. was (Author: gol...@detego-software.de): Please find attatched an extended TestSpanCollection.java that shows the problem. > Nested Span Queries are buggy > - > > Key: LUCENE-7398 > URL: https://issues.apache.org/jira/browse/LUCENE-7398 > Project: Lucene - Core > Issue Type: Bug > Components: core/search >Affects Versions: 5.5, 6.x >Reporter: Christoph Goller >Priority: Critical > Attachments: TestSpanCollection.java > > > Example for a nested SpanQuery that is not working: > Document: Human Genome Organization , HUGO , is trying to coordinate gene > mapping research worldwide. > Query: spanNear([body:coordinate, spanOr([spanNear([body:gene, body:mapping], > 0, true), body:gene]), body:research], 0, true) > The query should match "coordinate gene mapping research" as well as > "coordinate gene research". It does not match "coordinate gene mapping > research" with Lucene 5.5 or 6.1, it did however match with Lucene 4.10.4. It > probably stopped working with the changes on SpanQueries in 5.3. I will > attach a unit test that shows the problem. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Updated] (LUCENE-7398) Nested Span Queries are buggy
[ https://issues.apache.org/jira/browse/LUCENE-7398?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Christoph Goller updated LUCENE-7398: - Attachment: TestSpanCollection.java Please find attatched an extended TestSpanCollection.java that shows the problem. > Nested Span Queries are buggy > - > > Key: LUCENE-7398 > URL: https://issues.apache.org/jira/browse/LUCENE-7398 > Project: Lucene - Core > Issue Type: Bug > Components: core/search >Affects Versions: 5.5, 6.x >Reporter: Christoph Goller >Priority: Critical > Attachments: TestSpanCollection.java > > > Example for a nested SpanQuery that is not working: > Document: Human Genome Organization , HUGO , is trying to coordinate gene > mapping research worldwide. > Query: spanNear([body:coordinate, spanOr([spanNear([body:gene, body:mapping], > 0, true), body:gene]), body:research], 0, true) > The query should match "coordinate gene mapping research" as well as > "coordinate gene research". It does not match "coordinate gene mapping > research" with Lucene 5.5 or 6.1, it did however match with Lucene 4.10.4. It > probably stopped working with the changes on SpanQueries in 5.3. I will > attach a unit test that shows the problem. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] [Created] (LUCENE-7398) Nested Span Queries are buggy
Christoph Goller created LUCENE-7398: Summary: Nested Span Queries are buggy Key: LUCENE-7398 URL: https://issues.apache.org/jira/browse/LUCENE-7398 Project: Lucene - Core Issue Type: Bug Components: core/search Affects Versions: 5.5, 6.x Reporter: Christoph Goller Priority: Critical Example for a nested SpanQuery that is not working: Document: Human Genome Organization , HUGO , is trying to coordinate gene mapping research worldwide. Query: spanNear([body:coordinate, spanOr([spanNear([body:gene, body:mapping], 0, true), body:gene]), body:research], 0, true) The query should match "coordinate gene mapping research" as well as "coordinate gene research". It does not match "coordinate gene mapping research" with Lucene 5.5 or 6.1, it did however match with Lucene 4.10.4. It probably stopped working with the changes on SpanQueries in 5.3. I will attach a unit test that shows the problem. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] Closed: (LUCENE-2783) Deadlock in IndexWriter
[ https://issues.apache.org/jira/browse/LUCENE-2783?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ] Christoph Goller closed LUCENE-2783. Resolution: Fixed Already fixed with introduction of mergeDone flag in OneMerge of Lucene upcoming 2.9.4 > Deadlock in IndexWriter > --- > > Key: LUCENE-2783 > URL: https://issues.apache.org/jira/browse/LUCENE-2783 > Project: Lucene - Java > Issue Type: Bug > Components: Index >Affects Versions: 2.9.3 > Environment: ALL >Reporter: Christoph Goller > Fix For: 2.9.4 > > Original Estimate: 2h > Remaining Estimate: 2h > > If autoCommit == true a merge usually triggers a commit. A commit > (prepareCommit) can trigger a merge vi the flush method. There is a > synchronization mechanism for commit (commitLock) and a separate > synchronization mechanism for merging (ConcurrentMergeScheduler.wait). If one > thread holds the commitLock monitor and another one holds the > ConcurrentMergeScheduler monitor we have a deadlock. -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online. - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
[jira] Created: (LUCENE-2783) Deadlock in IndexWriter
Deadlock in IndexWriter --- Key: LUCENE-2783 URL: https://issues.apache.org/jira/browse/LUCENE-2783 Project: Lucene - Java Issue Type: Bug Components: Index Affects Versions: 2.9.3 Environment: ALL Reporter: Christoph Goller Fix For: 2.9.4 If autoCommit == true a merge usually triggers a commit. A commit (prepareCommit) can trigger a merge vi the flush method. There is a synchronization mechanism for commit (commitLock) and a separate synchronization mechanism for merging (ConcurrentMergeScheduler.wait). If one thread holds the commitLock monitor and another one holds the ConcurrentMergeScheduler monitor we have a deadlock. -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online. - To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org