![Page 1: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/1.jpg)
Sumit Bhatia
![Page 2: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/2.jpg)
Introduction
Inference Network based Retrieval Model
Experimental Results
Conclusions
2
![Page 3: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/3.jpg)
What are the important characteristics of aretrieval model?
◦ Document Representation
◦ Query Representation
◦ Matching Function
3
![Page 4: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/4.jpg)
Probabilistic Ranking Principle◦ P(R=1|d,q)
◦ 1-0 Loss model, BM25 Model
Language Models◦ P(d|q)
◦ Query Likelihood
◦ Relevance Based LM
◦ Translation Models
◦ Query document divergence
4
![Page 5: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/5.jpg)
Utilize multiple document representation schemes
5
CiteSeerX Document
TitleFull TextAbstract
![Page 6: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/6.jpg)
Information Need : To find about trade relationshipsbetween India and China
How to combine these results???
6
India China TradeSino India
Trade
![Page 7: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/7.jpg)
Nittany Lions vs Buckeyes
Or
Penn State vs Ohio State
Different vocabularies of document authorand query author
Mapping between related concepts
7
![Page 8: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/8.jpg)
Turtle and Croft (1991) introduced a retrievalmodel based on Inference Networks
Described IR as an inference process
◦ Users has information need I
◦ Expressed through a query Q
◦ A document is “observed”
◦ Estimate the probability that user’s informationneed is satisfied by the document
◦ Different query/document representations providethe evidence
8
![Page 9: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/9.jpg)
Inference Network (Inet)
• Directed Acyclic Graph
• Each node represents an event with a continuous/discrete outcome• In this talk, we assume binary outputs
• Root node event occurs with a prior probability
• Non root nodes store a table of conditionalprobabilities aka Link Matrix describing itsoutcome given parent node
9
![Page 10: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/10.jpg)
4 types of nodes:
◦ Document Nodes
◦ Representation Nodes
◦ Query Representation Nodes
◦ Information Need Node
10
![Page 11: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/11.jpg)
One node for each document in the corpus
Corresponds to the event that document di is observed
Only one document observed at a point
11
d1 d2 dn
![Page 12: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/12.jpg)
A set of representation nodes
Representation = an indexable feature of document, eg. terms, phrases etc.
Each document node acts as a parent for representation nodes
12
r1 r2 rk
![Page 13: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/13.jpg)
Since only one document is observed at a time, we wish to estimate the following for representation nodes:
Also known as belief, bel(rk)
13
r1 r2 rk
![Page 14: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/14.jpg)
Combine beliefs about representation nodes and other query nodes
Enable structured query operators
Need to compute Link Matrices
14
![Page 15: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/15.jpg)
15
Suppose query node q has n parent nodes with corresponding beliefs p1, p2, …, pn
There are 2n possible parent states and 2 possible outcomes of q
Represented by a 2 x 2n Link Matrix
![Page 16: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/16.jpg)
Represents the event that the information need of user is met
Combines evidence from all other query nodes into a single belief
Documents are ranked by this belief score
16
![Page 17: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/17.jpg)
17
D
R3R2R1
I
q2q1
Document Network
Query Network
Representation Nodebeliefs can be computedusing LMs
![Page 18: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/18.jpg)
Trec 4,6,7, and 8 datasets
For Trec 4, structured queries were available
For Trec 6,7 and 8 structured queries were generated from title and description fields
Example structured query for free text query “capital punishment deterrent crime”
18
![Page 19: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/19.jpg)
19
![Page 20: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/20.jpg)
Part of Lemur Toolkit
Based on Inference Networks and LMs
Powerful query language to support structured queries and evidence combination
Different document types – TREC style, XML, HTML, .doc
XML retrieval, sentence retrieval, passage retrieval, extent retrieval
Scalable
20
![Page 21: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/21.jpg)
Information Retrieval can be considered as aninference process
Inference Networks provide a theoreticallysound framework to infer the probability thata given document satisfies user’s informationneeds
Language Models provide a means toestimate the beliefs/probabilities
21
![Page 22: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/22.jpg)
1. D. Metzler and W.B. Croft. Combining the languagemodel and inference network approaches to retrieval.Info. Proc. and Mgt. 40(5):735-750, 2004.
2. Lemur Project Website: http://www.lemurproject.org
3. Trevor Strohman, Donald Metzler, Howard Turtle andW. Bruce Croft. Indri: A language-model based searchengine for complex queries. Umass Technical Report
4. H. Turtle and W. B. Croft, Evaluation of an inferencenetwork based retrieval model. Trans. Inf. Syst.,9(3):187-22, 1991.
22
![Page 23: Sumit Bhatiasumitbhatia.net/Talks/inet.pdf · 1. D. Metzler and W.B. Croft. Combining the language model and inference network approaches to retrieval. Info. Proc. and Mgt. 40(5):735-750,](https://reader033.vdocuments.net/reader033/viewer/2022050304/5f6c98f6110444796e051b03/html5/thumbnails/23.jpg)