Sure, here are some real world examples from my time at Netflix. Is this movie twice as much about “new york”?
* New York, New York Which one of these is the best match for “blade runner”: * Blade Runner: The Final Cut * Blade Runner: Theatrical & Director’s Cut * Blade Runner: Workprint http://dvd.netflix.com/Search?v1=blade+runner <http://dvd.netflix.com/Search?v1=blade+runner> At Netflix (when I was there), those were shown in popularity order with a boost function. And for stemming, should the movie “Saw” match “see”? Maybe not. wunder Walter Underwood wun...@wunderwood.org http://observer.wunderwood.org/ (my blog) > On Apr 20, 2016, at 5:28 PM, Jack Krupansky <jack.krupan...@gmail.com> wrote: > > Maybe it's a cultural difference, but I can't imagine why on a query for > "John", any of those titles would be treated as anything other than equals > - namely, that they are all about John. Maybe the issue is that this seems > like a contrived example, and I'm asking for a realistic example. Or, maybe > you have some rule of relevance that you haven't yet shared - and I mean > rule that a user would comprehend and consider valuable, not simply a > mechanical rule. > > > > -- Jack Krupansky > > On Wed, Apr 20, 2016 at 8:10 PM, <jimi.hulleg...@svensktnaringsliv.se> > wrote: > >> Ok sure, I can try and give some examples :) >> >> Lets say that we have the following documents: >> >> Id: 1 >> Title: John Doe >> >> Id: 2 >> Title: John Doe Jr. >> >> Id: 3 >> Title: John Lennon: The Life >> >> Id: 4 >> Title: John Thompson's Modern Course for the Piano: First Grade Book >> >> Id: 5 >> Title: I Rode With Stonewall: Being Chiefly The War Experiences of the >> Youngest Member of Jackson's Staff from John Brown's Raid to the Hanging of >> Mrs. Surratt >> >> >> And in general, when a search word matches the title, I would like to have >> the length of the title field influence the score, so that matching >> documents with shorter title get a higher score than documents with longer >> title, all else considered equal. >> >> So, when a user searches for "John", I would like the results to be pretty >> much in the order presented above. Though, it is not crucial that for >> example document 1 comes before document 2. But I would surely want >> document 1-3 to come before document 4 and 5. >> >> In my mind, the fieldNorm is a perfect solution for this. At least in >> theory. In practice, the encoding of the fieldNorm seems to make this >> function much less useful for this use case. Unless I have missed something. >> >> Is there another way to achive something like this? Note that I don't want >> a general boost on documents with short titles, I only want to boost them >> if the title field actually matched the query. >> >> /Jimi >> >> ________________________________________ >> From: Jack Krupansky <jack.krupan...@gmail.com> >> Sent: Thursday, April 21, 2016 1:28 AM >> To: solr-user@lucene.apache.org >> Subject: Re: Is it possible to configure a minimum field length for the >> fieldNorm value? >> >> I'm not sure I fully follow what distinction you're trying to focus on. I >> mean, traditionally length normalization has simply tried to distinguish a >> title field (rarely more than a dozen words) from a full body of text, or >> maybe an abstract, not things like exactly how many words were in a title. >> Or, as another example, a short newswire article of a few paragraphs vs. a >> feature-length article, paper, or even book. IOW, traditionally it was more >> of a boolean than a broad range of values. Sure, yes, you absolutely can >> define a custom similarity with a custom norm that supports a wide range of >> lengths, but you'll have to decide what you really want to achieve to tune >> it. >> >> Maybe you could give a couple examples of field values that you feel should >> be scored differently based on length. >> >> -- Jack Krupansky >> >> On Wed, Apr 20, 2016 at 7:17 PM, <jimi.hulleg...@svensktnaringsliv.se> >> wrote: >> >>> I am talking about the title field. And for the title field, a sweetspot >>> interval of 1 to 50 makes very little sense. I want to have a fieldNorm >>> value that differentiates between for example 2, 3, 4 and 5 terms in the >>> title, but only very little. >>> >>> The 20% number I got by simply calculating the difference in the title >>> fieldNorm of two documents, where one title was one word longer than the >>> other title. And one fieldNorm value was 20% larger then the other as a >>> result of that. And since we use multiplicative scoring calculation, a >> 20% >>> increase in the fieldNorm results in a 20% increase in the final score. >>> >>> I'm not talking about "scores as percentages". I'm simply noting that >> this >>> minor change in the text data (adding or removing one single word) causes >>> the score to change by a almost 20%. I noted this when I renamed a >>> document, removing a word from the title, and that single change caused >> the >>> document to move up several positions in the result list. We don't want >>> such minor modifications to have such big impact of the resulting score. >>> >>> I'm not sure I can agree with you that "the effect of document length >>> normalization factor is minimal". Then why does it inpact our result in >>> such a big way? And as I said, we don't want to disable it completely, we >>> just want it to have a much lesser effect, even on really short texts. >>> >>> /Jimi >>> >>> ________________________________________ >>> From: Ahmet Arslan <iori...@yahoo.com.INVALID> >>> Sent: Thursday, April 21, 2016 12:10 AM >>> To: solr-user@lucene.apache.org >>> Subject: Re: Is it possible to configure a minimum field length for the >>> fieldNorm value? >>> >>> Hi Jimi, >>> >>> Please define a meaningful document-lenght range like min=1 max=50. >>> By the way you need to reindex every time you change something. >>> >>> Regarding 20% score change, I am not sure how you calculated that number >>> and I assume it is correct. >>> What really matters is the relative order of documents. It doesn't mean >>> anything addition of a word decreases the initial score by x%. Please >> see : >>> https://wiki.apache.org/lucene-java/ScoresAsPercentages >>> >>> There is an information retrieval heuristic which says that addition of a >>> non-query term should decrease the score. >>> >>> Lucene's default document length normalization may favor short document >>> too much. But folks blend score with other structural fields >> (popularity), >>> even completely bypass relevancy score and order by price, production >> date >>> etc. I mean there are many use cases, the effect of document length >>> normalization factor is minimal. >>> >>> Lucene/Solr is highly pluggable, very easy to customize. >>> >>> Ahmet >>> >>> >>> On Wednesday, April 20, 2016 11:05 PM, " >>> jimi.hulleg...@svensktnaringsliv.se" < >> jimi.hulleg...@svensktnaringsliv.se> >>> wrote: >>> Hi Ahmet, >>> >>> SweetSpotSimilarity seems quite nice. Some simple testing by throwing >> some >>> different values at the class gives quite good results. Setting ln_min=1, >>> ln_max=2, steepness=0.1 and discountOverlaps=true should give me more or >>> less what I want. At least for the title field. I'm not sure what the >>> actual effect of those settings would be on longer text fields, so maybe >> I >>> will use the SweetSpotSimilarity only for the title field to start with. >>> >>> Of course I understand that there are many things that can be considered >>> domain specific requirements, like if to favor/punish short/medium/long >>> texts, and how. I was just wondering how many actual use cases there are >>> where one want's a ~20% difference in score between two documents, where >>> the only difference is that one of the documents has one extra word in >> one >>> field. (And now I'm talking about an extra word that doesn't affect >>> anything else except the fieldNorm value). I for one find it hard to find >>> such a use case, and would consider it a very special use case, and would >>> consider a more lenient calculation a better fit for most use cases (and >>> therefore most domains). :) >>> >>> /Jimi >>> >>> >>> -----Original Message----- >>> From: Ahmet Arslan [mailto:iori...@yahoo.com.INVALID] >>> Sent: Wednesday, April 20, 2016 8:14 PM >>> To: solr-user@lucene.apache.org >>> Subject: Re: Is it possible to configure a minimum field length for the >>> fieldNorm value? >>> >>> Hi Jimi, >>> >>> SweetSpotSimilarity allows you define a document length range, so that >> all >>> documents in that range will get same fieldNorm value. >>> In your case, you can say that from 1 word up to 100 words do not employ >>> document length punishment. If a document is longer than 100 do some >>> punishment. >>> >>> By the way; favoring/punishing short, middle, or long documents is >> domain >>> specific thing. You are free to decide what to do. >>> >>> Ahmet >>> >>> >>> >>> On Wednesday, April 20, 2016 7:46 PM, " >> jimi.hulleg...@svensktnaringsliv.se" >>> <jimi.hulleg...@svensktnaringsliv.se> wrote: >>> OK. Well, still, the fact that the score increases almost 20% because of >>> just one extra term in the field, is not really reasonable if you ask me. >>> But you seem to say that this is expected, reasonable and wanted behavior >>> for most use case? >>> >>> I'm not sure that I feel comfortable replacing the default Similarity >>> implementation with a custom one. That would just increase the complexity >>> of our setup and would make future upgrades harder (we would for example >>> have to remember to check if the default similarity configuration or >>> implementation changes). >>> >>> No, if it really is the case that most people like and want this, and >>> there is no way to configure Solr/Lucene to calculate fieldNorm in a more >>> reasonable way (in my book) for short field values, then I just think we >>> are forced to set omitNorms="true", maybe in combination with a simple >>> field boost for shorter fields. >>> >>> /Jimi >>> >>> >>> >>> -----Original Message----- >>> From: Jack Krupansky [mailto:jack.krupan...@gmail.com] >>> Sent: Wednesday, April 20, 2016 5:18 PM >>> To: solr-user@lucene.apache.org >>> Subject: Re: Is it possible to configure a minimum field length for the >>> fieldNorm value? >>> >>> FWIW, length for normalization is measured in terms (tokens), not >>> characters. >>> >>> With TDIFS similarity (the default before 6.0), the normalization is >> based >>> on the inverse square root of the number of terms in the field: >>> >>> return state.getBoost() * ((float) (1.0 / Math.sqrt(numTerms))); >>> >>> That code is in ClassicSimilarity: >>> >>> >> https://github.com/apache/lucene-solr/blob/releases/lucene-solr/5.5.0/lucene/core/src/java/org/apache/lucene/search/similarities/ClassicSimilarity.java#L115 >>> >>> You can always write your own custom Similarity class to override that >>> calculation. >>> >>> -- Jack Krupansky >>> >>> On Wed, Apr 20, 2016 at 10:43 AM, <jimi.hulleg...@svensktnaringsliv.se> >>> wrote: >>> >>>> Hi, >>>> >>>> In general I think that the fieldNorm factor in the score calculation >>>> is quite good. But when the text is short I think that the effect is >> two >>> big. >>>> >>>> Ie with two documents that have a short text in the same field, just a >>>> few characters extra in of the documents lower the fieldNorm factor too >>> much. >>>> In one test the text in document 1 is 30 characters long and has >>>> fieldNorm 0.4375, and in document 2 the text is 37 characters long and >>>> has fieldNorm 0.375. That means that the first document gets almost a >>>> 20% higher score simply because of the 7 character difference. >>>> >>>> What are my options if I want to change this behavior? Can I set a >>>> lower character limit, meaning that all fields with a length below >>>> this limit gets the same fieldNorm value? >>>> >>>> I know I can force fieldNorm to be 1 by setting omitNorms="true" for >>>> that field, but I would prefer to still have it, just limit its effect >>>> on short texts. >>>> >>>> Regards >>>> /Jimi >>>> >>>> >>>> >>> >>
