If you think this page looks a lot like the
PivotedDocumentLengthNormalization, you're right. I used Jinyoung's format. Thanks Jinyoung!
| Date | Place | Author | Keyword(s) |
| 2001 | SIGIR | Laverenko & Croft | relevance model |
Summary
Introduction of a modification to the language model framework using the notion of a relevance model: probabilities of words in the relevant class. This work showed how this relevance estimation can be done using nothing more than queries.
Background
Language Models
- Statistical analysis of corpus to determine probability of P(Q|D) and likelihood of P(D|Q)
- Originally does not allow for features not based explicitly on statistical properties of corpora/queries
Vector-Space Model
- Much less constrained, allows for arbitrary notions of similarity/distance
- Lacks much empirical backing - usually based on heuristics
Contribution
Relevance (Topic) Model
- Instead of calculating P(w|D), we calculate P(w|R) - the relevance of a word to a particular information need (modeled)
Comment
- Relevance Models allow for a higher level of feature expression
Reference
to top