On 09/03/2013 07:33 PM, David Miranda wrote:
Is there any way to check the similarity of texts with Lucene? I have the
DBpedia indexed and wanted to get the texts more similar between the
abstract and DBpedia another text. If I do a search in the abstract field,
with a particular text the result is not very satisfactory. Eg Abstract
DBpedia: "SoundCloud is an online audio distribution platform Which Allows
collaboration, promotion and distribution of audio recordings." My Text:
"Private Track From DJ Sneak. Download the track now in the SoundCloud
website."
You are attacking extremly hard problem here -- searching short documents
with a long query. This creates a lots of problems, as setting document
frequency of a term to the same magnitude of its own frequency which
instantly kills some similarity measures.
All you can do is to experiment a lot with different similarity measures
and preprocessing steps.
Sim measures are simple, just try them all for each preprocessing
combination.
Suggestions of preprocessing steps:
- remove all stop words
- remove all functional words (you can find list of them at wikipedia)
- boost all uppercase words or words containing at least one uppercase
letter (add boost of 3 or 4; maybe skip first word of a sentence)
- break search text into sentences then search index for each sentence
(combine results using borda count or something similar)
- do what Koji suggested
Regards,
Ivan Krišto
