Download presentation
Presentation is loading. Please wait.
1
11 September 2002IR/LM workshop, Amherst1 Information retrieval, language and ‘language models’ Stephen Robertson Microsoft Research Cambridge and City University London
2
11 September 2002IR/LM workshop, Amherst2 Language and IR IR deals mainly in text objects Text = language Therefore, models or theories about language must be relevant to IR Many suggestions/attempts –Transformational methods –Shallow or deep NLP –Anaphora etc. etc.
3
11 September 2002IR/LM workshop, Amherst3 Language and IR But IR went its own sweet way –Term weighting, scoring functions, vector spaces, probabilistic models… –… with a strong emphasis on statistics Eventually, the language people became interested in statistics –Statistical NLP, collocation linguistics…
4
11 September 2002IR/LM workshop, Amherst4 Language and IR But ‘language models’ (as in this workshop title) seem to come from outside … and to share with IR a cavalier view of language So, can language models succeed where other language approaches have failed?
5
11 September 2002IR/LM workshop, Amherst5 Some modelling issues Relevance Topicality Learning Sources of evidence
6
11 September 2002IR/LM workshop, Amherst6 Relevance Central question: what is good system behaviour (what does the user want to see? what would satisfy him/her) Not necessarily a binary Relevance variable, though that has proved very useful Early language models seemed to hide this –but this is changing
7
11 September 2002IR/LM workshop, Amherst7 Topicality How do we understand ‘topics’? Documents are multi-topic Topics are not predefined… … potentially, any query defines a new topic (or perhaps more than one?) Models of topicality have eluded the IR community… … thus providing a significant opportunity for language modelling approaches
8
11 September 2002IR/LM workshop, Amherst8 Learning and Sources of evidence The major question: how to learn… … and from what? E.g. classical relevance feedback Text of query… … + relevance judgements So how do we combine this evidence? Again, opportunities for language models
9
11 September 2002IR/LM workshop, Amherst9 Final remarks Information retrieval is a slippery domain for modelling Language modelling has the potential to add significantly to the modelling tools available There are many connections between modelling approaches that need exploring
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.