Hi Robert, this is the kind of scenario I have worked on in the last couple of years in my previous company. Adding semantic and natural language capabilities to your indexing pipeline could help a lot. First of all you need a meaningful knowledge base describing your business ontology. i.e. having the ground knowledge to let your machine understand that an Ipad is a Tablet, and not a pill but an electronical device. Adding a Named Entity Linking layer to your indexing pipeline ( with the configured knowledge base) you can first of all identify at indexing level when an occurrence of an entity should be linked to a real world object. In your case, assuming a meaningful knowledge base, Ipad occurrences will be linked to the Ipad entity which is of type Tablet which is of type Electronical Device. At this point you need to model your index with nested object and manage the query time side. Of course it is not an immediate solution, but the benefit could be good and you can get closer to natural language search.
Take a look to my Lucene Revolution Presentation and some old blogs : https://lucidworks.com/blog/2015/08/31/apache-solr-multi-language-content-discovery-entity-driven-search/ http://www.zaizi.com/blog/sensefy-content-discovery-through-entity-driven-search On 9 March 2016 at 10:53, Charlie Hull <char...@flax.co.uk> wrote: > On 09/03/2016 10:05, Robert Brown wrote: > >> Hi, >> >> I'm looking for some advice and possible options for dealing with our >> relevancy when searching through shopping products. >> >> A search for "tablet" returns pills, when the user would expect >> electronic devices. >> >> Without any extra criteria (like category), how would/could you manage >> this situation? >> >> Any solution would also need to scale since this is just a random example. >> >> Thanks, >> Rob >> >> Hi Rob, > > Solr out of the box has no way of knowing that 'the user would expect > electronic devices', unfortunately: since the record for 'pills' contains > the word 'tablet', that's what you get. Note that if your users were > expecting a medical answer everything would be rosy! > > Firstly, consider setting up some tests for these kinds of issues, so you > can measure if the adjustments you're making are having the effect you want > - we call this test-driven relevancy tuning. One tool you might consider > (if you don't want to use a pile of spreadsheets) is Quepid (disclaimer: we > resell this in the UK). > > Then, you need to work out *why* the results are wrong for your use case > (using Solr's debugQuery helps here). If the words in the body text are > having a disproportionate effect, consider boosting another part of the > source data. Consider synonyms (if I search 'tablet' I should also get > 'iPad'). To be honest this is a complex field with a lot of different knobs > to adjust - I would recommend you take a look at Doug Turnbull and John > Berryman's new book 'Relevant Search' (available on MEAP at Manning > Publications) which is an excellent take on this. > > In short, you need a sensible methodology for tuning relevance, otherwise > it can easily become a game of whack-a-mole! > > Cheers > > Charlie > > -- > Charlie Hull > Flax - Open Source Enterprise Search > > tel/fax: +44 (0)8700 118334 > mobile: +44 (0)7767 825828 > web: www.flax.co.uk > -- -------------------------- Benedetti Alessandro Visiting card : http://about.me/alessandro_benedetti "Tyger, tyger burning bright In the forests of the night, What immortal hand or eye Could frame thy fearful symmetry?" William Blake - Songs of Experience -1794 England
