Book: Bayesian Networks : A practical guide to applications Paper-authors: Luis M. de Campos, Juan M. Fernandez-Luna, Juan F. Huete, Carlos Martine, Alfonso E. Romero Chapter: 12 Presented by Quratulain CSE 655 Probabilistic Reasoning Faculty of Computer Science, Institute of Business Administration
Outline Introduction Overview of information retrieval systems Bayesian network and information retrieval Theoretical foundations Building the information retrieval system Conclusion 10 oct, 20092Quratulain
Introduction/Motivation To fulfil the objective of democracy, need to make public all activities of parliament. Previously, information was sent in a printed form to all official organization and libraries. Currently, electronic document published on the web, which is fast, cheaper and an easier way. The official bulletin, transcripts of all speeches in different session, after editing published on website in PDF. The documents are accessible using database-like queries. 10 oct, 20093Quratulain
Problems To access information user must know about: Session number Date of legislature Difficult to access information 10 oct, 2009Quratulain4
Goal A website with real search engine based on content. The natural language query is applied to access the information. The obtained the relevant document through system. The output will be a set of document components of varying granularity (from complete document to single paragraph, also sorted depending on degree of relevance). ** This will avoid manual search ** 10 oct, 2009Quratulain5
Outline Introduction Overview of information retrieval systems Bayesian network and information retrieval Theoretical foundations Building the information retrieval system Conclusion 10 oct, 20096Quratulain
Overview of information retrieval Information retrieval is concerned with representation, storage, organization, and accessing of information items. Information retrieval systems work as: Given a set of documents Pre-processing remove words not useful in search(stopwords) Convert word to its stem word(reduce vocabulary) Each word is associated with weights expressing their importance (in document or collection of documents) NLP query indexed to match query representation with the stored document using any IR model. Finally, a set of document identifiers is presented to the user sorted according to their relevance degree. 10 oct, 2009Quratulain7
Overview of information retrieval Standard IR treat document as atomic entities. XML allows structured documents with semantics. Structured IR views documents as aggregates interrelated structural elements by indexing. Structured IR models exploit the content and the structure of documents to estimate the relevance of document components to query. 10 oct, 2009Quratulain8
Outline Introduction Overview of information retrieval systems Bayesian network and information retrieval Theoretical foundations Building the information retrieval system Conclusion 10 oct, 20099Quratulain
Bayesian Networks and information retrieval Bayesian networks were first applied to IR at the beginning of 1990 by croft and turtle. Bayesian network in IR models compute the probability of relevance given a document and a query. Two important model of BNs within IR: Belief network model Bayesian network retrieval model. Common feature are: Each index term and document represented as nodes in network. Links connecting each document node with all the term nodes. Model differ in: The direction of arc. Additional arc (relationship b/w documents and terms.) 10 oct, 2009Quratulain10
BN-based retrieval model 10 oct, 2009Quratulain11 D2 T1 D1 T7T6T5T4T3T2 D3 Terms Documents
Drawback of Bayesian network 1. Time and space require to assess the distributions and store them (conditional probability per node is exponential with the parent nodes ) 2. The efficiency of carrying out inference, because general inference in BNs is NP-hard problem Therefore The direct approach where we propagate the evidence contained in a query through the whole network is unfeasible. 10 oct, 2009Quratulain12
Outline Introduction Overview of information retrieval systems Bayesian network and information retrieval Theoretical foundations Building the information retrieval system Conclusion 10 oct, Quratulain
Theoretical foundations Set of documents D={D 1,D 2,..., D M } Set of terms used to index these documents Each document D i is organized hierarchically, representing structural associations of elements in D i called structural unit. These association to a document form a tree. For example scientific article. 10 oct, 2009 Quratulain 14
The structure of scientific article 10 oct, 2009Quratulain15 Index Terms TitleParag 1Parag 2TitleParag 1 TitleParag 1 Ref 1Ref 2 Subsec 1Subsec 2 Section 1 Section 2 Bibligraphy TitleAuthorAbstract Document 1
BN model for document BN modeling of document contain 3-kind of nodes Terms set, T={T 1, T 2,..., T l } Basic structural unit, U b ={B 1, B 2,..., B m } Complex structural unit, U c ={S 1, S 2,..., S m } Set of all structural unit U= U b U c To each node T, B, S is associated a binary random variables as {t -, t + }, {b -, b + } or {s -, s + } respectively. (-) not relevant, (+) relevant. 10 oct, 2009Quratulain16
BN model for document 10 oct, 2009Quratulain17 UbUb T1T6T11T10T9T8T2T3T4T5T7T16T15T14T13T12B1B6B2B3 B4 B5B7S1 S2 S3 S4 UcUc UcUc UsUs, with Pa(S 1 )Pa(S 2 ) =,S1S1 S2S2 UcUc
BN for document Conditional Probability P(t + ) P(b + |pa(B)) P(s + |pa(S)) Due to greater number of parent, efficient inference procedure is needed. 10 oct, 2009Quratulain18
Influence Diagram Model Once the BN has been constructed transform it into influence diagram by including decision and utility nodes. Chance node : previous BN Decision node : Utility node : 10 oct, 2009Quratulain19
Outline Introduction Overview of information retrieval systems Bayesian network and information retrieval Theoretical foundations Building the information retrieval system Conclusion 10 oct, Quratulain
Building the information retrieval system(PAIRS) PAIRS is a software package (store document in relational database) Written in C++ Specifically developed to store and retrieve documents generated by the parliament of Andalusia Based on probabilistic model. 10 oct, 2009Quratulain21 PDF document collection XML document collection Indexing System Query Indexed Query Search Engine Indexed Document Collection Retrieved Document Components General scheme of PAIRS
Outline Introduction Overview of information retrieval systems Bayesian network and information retrieval Theoretical foundations Building the information retrieval system Conclusion 10 oct, Quratulain
Conclusion This paper present a retrieval system based on probabilistic model belong to parliament information. The system has been proven efficient in term of indexing and retrieval time. Bayesian network technologies can be employed in problem domains whose dimensionality would earlier avoid its use. The system is not a finished product, still several possible improvement are required. 10 oct, 2009Quratulain23