David, Thank you for taking the time to evaluate SOLR-2585. Perhaps the title of the issue advertises more than it delivers? (The name is borrowed from a section in the first book listed here: http://wiki.apache.org/lucene-java/InformationRetrieval) In any case, I think SOLR-2585 is a step forward. The idea is that some words are "correctly spelled" in that they exist in the dictionary, yet are incorrect in the context of the user's query. The patch that is out there just tries to find the user something that works. It sounds like you want it to find the _best_ something that works, and its not doing a good job at that.
The solution so far is crude: it just takes the most promising words (based on low edit-distance and higher doc frequency) and tries re-querying different combinations until it finds some that give you hits. There are no doubt a ton of ways to make this more efficient (and effective). The book I mention says to look at 2-word shingle combinations or possibly check the query log for combinations that have worked in the past. I would imagine as time goes on someone would implement things like this for Solr. Your idea to have it consider term proximity is interesting. Perhaps we can hack this with the current code by changing your "spellcheck.q" to a phrase query? Or if the user had more than 2 words adding slop as well so that it would consider the words 1-or-2 removed but not further? Of course this will *eliminate* collations that don't meet the phrase requirememnts and you probably would rather just have it rank them lower, right? (this would require better code!) In my use-cases we usually require 100% terms (mm=100%), so at the time "spellcheck.maxResultsForSuggest" seemed to make sense. If all the terms are required, then by default the spellchecker returns nothing if even only 1 result is returned. So from my perspective this parameter makes it more flexible: set it to 5, for instance, and now you get spelling suggestions if the query returns 0-5 results. But I can see how this might not be what you want in cases where mm<100%. It might be awful for mm=0. Can you think of something better? One more thing: Would you be ok if I took some of our comments here and added them to the JIRA issue? As this is code that is not even in trunk, it would be helpful to track our comments in JIRA and get visibility on the dev-list also, where discussions about unincorporated patches usually occur. James Dyer E-Commerce Systems Ingram Content Group (615) 213-4311 -----Original Message----- From: David Radunz [mailto:da...@boxen.net] Sent: Sunday, January 22, 2012 6:42 AM To: solr-user@lucene.apache.org Subject: Re: Improving Solr Spell Checker Results Hey James, I have played around a bit more with the settings and tried setting spellcheck.maxResultsForSuggest=100 and spellcheck.maxCollations=3. This yields 'Sigourney Weaver' as ONE of the corrections, but it's the second one and not the first. Which is wrong if this is a patch for 'context sensative', because it doesn't really seem to honor any context at all. Unless I am missunderstanding this? Also, I don't really like maxResultsForSuggest as it means 'all or nothing'. If you set it to 10 and there are 100 results, then you offer no corrections at all even if the term is missing in the dictionary entirely. If I set spellcheck.maxResultsForSuggest=100 and spellcheck.maxCollations=3 and choose the collation with the largest 'hits' I get Sigourney Weaver and other 'popular' terms. But say I searched for 'pork and chups', the 'popular' correction is 'park and chips' where as the first correction was correct: 'pork and chips'. So really, none of the solutions either in this patch or Solr offer what I would truely call context sensative spell checking. That being, in a full text search engine you find documents based on terms and how close they are togehter in the document. It makes more than perfect sense to treat the dictionary like this, so that when there are multiple terms it offers suggestions for the terms that match closely to whats entered surrounding the term. Example: "Sigourney Wever" would never appear in a document ever. "Sigourney Weaver" however has many 'hits' in exactly that order of words. So there needs to be a way to boost suggestions based on adjacency... Much like the full text search operates. Thoughts? David On 22/01/2012 9:56 PM, David Radunz wrote: > James, > > I worked out that I actually needed to 'apply' patch SOLR-2585, > whoops. So I have done that now and it seems to return > 'correctlySpelled=true' for 'Sigorney Wever' (when Sigorney isn't even > in the dictionary). Could something have changed in the trunk to make > your patch no longer work? I had to manually merge the setup for the > test case due to a new 'hyphens' test case. The settings I am use are: > > <lst name="defaults"> > <str name="echoParams">explicit</str> > <int name="rows">10</int> > > <str name="spellcheck.onlyMorePopular">false</str> > <int name="spellcheck.count">10</int> > <str name="spellcheck.extendedResults">true</str> > <str name="spellcheck.collate">true</str> > <str name="spellcheck.collateExtendedResults">true</str> > <int name="spellcheck.maxCollationTries">10</int> > <int name="spellcheck.maxCollations">1</int> > > <int name="spellcheck.alternativeTermCount">5</int> > <int name="spellcheck.maxResultsForSuggest">1</int> > </lst> > > > <lst name="spellchecker"> > <str name="name">default</str> > <str name="field">spell</str> > <str name="classname">solr.DirectSolrSpellChecker</str> > > <!-- the spellcheck distance measure used, the default is the internal > levenshtein --> > <str name="distanceMeasure">internal</str> > <!-- minimum accuracy needed to be considered a valid spellcheck > suggestion --> > <float name="accuracy">0.5</float> > <!-- the maximum #edits we consider when enumerating terms: can be 1 > or 2 --> > <int name="maxEdits">2</int> > <!-- the minimum shared prefix when enumerating terms --> > <int name="minPrefix">1</int> > <!-- maximum number of inspections per result. --> > <int name="maxInspections">5</int> > <!-- minimum length of a query term to be considered for correction --> > <int name="minQueryLength">4</int> > <!-- maximum threshold of documents a query term can appear to be > considered for correction --> > <float name="maxQueryFrequency">0.01</float> > <!-- require suggestions to occur in 0.1% of the documents --> > <!-- > <float name="thresholdTokenFrequency">0.001</float> > --> > > <str name="spellcheckIndexDir">spellchecker</str> > <str name="buildOnCommit">true</str> > </lst> > > With the query: > > spellcheck=true&facet=on&fl=id,sku,name,format,thumbnail,release_date,url_path,price,special_price,year_made_attr_opt_combo,primary_cat_id&sort=score+desc,name+asc,year_made+desc&start=0&q=sigorney+wever+title:"sigorney+wever"^100+series_name:"sigorney+wever"^50&spellcheck.q=sigorney+wever&fq=store_id:"1"&rows=5 > > > > Cheers, > > David > > > On 22/01/2012 2:03 AM, David Radunz wrote: >> James, >> >> Thanks again for your lengthy and informative response. I updated >> from SVN trunk again today and was successfully able to run 'ant >> test'. So I proceeded with trying your suggestions (for question 1 so >> far): >> >> On 17/01/2012 5:32 AM, Dyer, James wrote: >>> David, >>> >>> The spellchecker normally won't give suggestions for any term in >>> your index. So even if "wever" is misspelled in context, if it >>> exists in the index the spell checker will not try correcting it. >>> There are 3 workarounds: >>> 1. Use the patch included with SOLR-2585 (this is for Trunk/4.x >>> only). See https://issues.apache.org/jira/browse/SOLR-2585 >> I have tried using this with the original test case of 'Signorney >> Wever'. I didn't notice any difference, although I am a little >> unclear as to what exactly this patch does. Nor am I really clear >> what to set either of the options to, so I set them both to '5'. I >> tried to find the test case it mentions, but it's not present in >> SpellCheckCollatorTest.java .. Any suggestions? >> >>> 2. try "onlyMorePopular=true" in your request. >>> (http://wiki.apache.org/solr/SpellCheckComponent#spellcheck.onlyMorePopular). >>> >>> But see the September 2, 2011 comment in SOLR-2585 about why this >>> might not do what you'd hope it would. >> >> Trying this did produce 'Signourney Weaver' as you would hope, but I >> am a little afraid of the downside. I would much more like a context >> sensative spell check that involves the terms around the correction. >>> >>> 3. If you're building your index on a<copyField />, you can add a >>> stopword filter that filters out all of the misspelt or rare words >>> from the field that the dictionary is based. This could be an >>> arduous task, and it may or may not work well for your data. >> I am currently using a copyField for all terms that are relevant, >> which is quite a lot and the dictionary would encompass a huge amount >> of data. Adding stopword filters would be out of the question as we >> presently have more than 30,000 products and this is for the initial >> launch, we intend to have many many more. >>> >>> As for your second question, I take it you're using (e)dismax with >>> multiple fields in "qf", right? The only way I know to handle this >>> is to create a<copyfield> that combines all of the fields you >>> search across. Use this combined field to base your dictionary. >>> Also, specifying "spellcheck.maxCollationTries" with a non-zero >>> value will weed out the nonsense word combinations that are likely >>> to occur when doing this, ensuring that any collations provided will >>> indeed yield hits. The downside to doing this, of course, is it >>> will make your first problem more acute in that there will be even >>> more terms in your index that the spellchecker will ignore entirely, >>> even if they're mispelled in context. Once again, SOLR-2585 is >>> designed to tackle this problem but it is still in its early stages, >>> and thus far it is Trunk-only. >> I tried setting spellcheck.maxCollationTries to 5 to see if it would >> help with the above problem, but it did not. >> >> I have now tried using it in the context of question 2. I tried >> searching for 'Sigorney Wever' in the series name (which it's not >> present in, as its an actor): >> >> spellcheck=true&facet=on&fl=id,sku,name,format,thumbnail,release_date,url_path,price,special_price,year_made_attr_opt_combo,series_name_attr_opt_combo&sort=score+desc,release_date+desc&start=0&q=*+series_name:"signourney+wever"^100&spellcheck.q=signourney+wever&fq=store_id:"1"+AND+series_name_attr_opt_search:*signourney*wever*&rows=5&spellcheck.maxCollationTries=5 >> >> >> >> Suggestions for 'Sigourney' Wever were returned, but no spelling >> suggestions or ones for series names (which i doubt there would be) >> should have been returned. >> >>> >>> You might also be interested in >>> https://issues.apache.org/jira/browse/SOLR-2993 . Although this is >>> unrelated to your two questions, the patch on this issue introduces >>> a new "ConjunctionSolrSpellChecker" which theoretically could be >>> enhanced to do exactly what you want. That is, you could >>> (theoretically) create separate dictionaries for each of the fields >>> you're searching and let the CSSC combine the results& generate >>> collations, etc. >> >> During the upgrade I switched to solr.DirectSolrSpellChecker, which I >> presume will help with this? I am a senior developer (in >> Java/Perl/Python/PHP) but I have not as yet looked at any of the Solr >> source code. So I am in the dark when you say it could be tailored >> for my needs. Also, how would it work? Query wise.. Would it be >> like.. spellcheck.series_name.q= and spellcheck.actor.q= and so on? >> If so that sounds tempting to try and achieve. But if you could >> provide any pointers in what exactly would be required that would >> really help. >> >> Thanks again for your time, >> >> David >>> >>> James Dyer >>> E-Commerce Systems >>> Ingram Content Group >>> (615) 213-4311 >>> >>> >>> -----Original Message----- >>> From: David Radunz [mailto:da...@boxen.net] >>> Sent: Friday, January 13, 2012 11:42 PM >>> To: solr-user@lucene.apache.org >>> Subject: Improving Solr Spell Checker Results >>> >>> Hey, >>> >>> Firstly I would like to thank you all for creating such a great >>> searching platform. What I was wondering is whether it is possible to: >>> >>> 1. Have the spell checker take into account multiple words. For example >>> if I search for "Sigourney Wever" it doesn't flag as a spelling >>> issue as >>> 'wever' is a correctly spelled word. And if I searched for "Sigourney >>> Wevr" the suggestion is "Sigourney Wever". Of course the correct >>> spelling is: Sigourney Weaver >>> 2. Have the spell checker return corrections only for dictionary items >>> added on the field being searched. i.e. Searching for an actor would >>> only use the dictionary fields from the actor. This makes sense on many >>> levels, as when you are field searching its useless to get a correction >>> from another field as no values would match in any case. >>> >>> Hopefully someone can help! >>> >>> Thanks in advance, >>> >>> David >> >
