Information Retrieval and Extraction 2010 Term Project – Modern Web Search Advisor: 陳信希 TA: 許名宏 & 王界人
Overview (in English) Goal Goal –Using advanced approaches to enhance Okapi-BM25 Group Group –1~3 person(s) per group; the name list to the TA Approach Approach –No limitations; Any resources on the Web is usable. Date of system demo and report submission Date of system demo and report submission –6/24 Thursday (provisional) Grading criteria Grading criteria –Originality and reasonableness of your approach –Effort for implementation / per person –Retrieval performance (training & testing) –Completeness of the report ( 分工、結果分析 )
Overview (in Chinese) 專題目標 專題目標 – 以進階 IR 技術提升 Okapi-BM25 的效能 分組 分組 –1~3 人 / 組,請組長將組員名單 ( 學號、姓名 ) 給 TA 方法 方法 – 不限,可使用任何 toolkit or resource on Web Demo 及報告繳交 Demo 及報告繳交 –6/25 Friday 評分標準 評分標準 – 所採用的方法創意、合理性 –Effort of implementation / per person – 檢索效能 (training 、 testing) – 報告完整性、分工及檢索結果分析
Content of Report Detail description about your approach Detail description about your approach Parameter setting (if parametric) Parameter setting (if parametric) System performance on the training topics System performance on the training topics –The baseline (Okapi-BM25) performance –The performance of your approach Division of the work ( 如何分工 ) Division of the work ( 如何分工 ) What you have learned ( 心得 ) What you have learned ( 心得 ) Others (optional) Others (optional)
Baseline Implementation: Okapi-BM25 Parametric probabilistic model Parametric probabilistic model Parameter setting Parameter setting –k 1 =1.2, k 2 =0, k 3 =0, b =0.75, R =r =0 (initial guess) Stemming: Porter ’ s stemmer Stemming: Porter ’ s stemmerPorter ’ s stemmerPorter ’ s stemmer
Possible Approaches Pseudo relevance feedback (PRF) Pseudo relevance feedback (PRF) –Supported by Lemur API Simple and effective, but no originality Simple and effective, but no originality Query expansion Query expansion –Using external resources ex: WordNet, Wikipedia, query log (AOL)...etc AOL Word sense disambiguation in docs/query Word sense disambiguation in docs/query Combining Results from 2 or more IR systems Combining Results from 2 or more IR systems Latent semantic analysis (LSI) Latent semantic analysis (LSI) Others Others –learning to rank, clustering/classification, …
Experimental Dataset A partial collection of TREC WT10g A partial collection of TREC WT10g –~10k documents –Link information is provided 30 topics for system development (training) 30 topics for system development (training) Another 20 topics in demo (testing) Another 20 topics in demo (testing)
Topic Example <top> Number: 476 Number: 476 Jennifer Aniston Jennifer Aniston Description: Description: Find documents that identify movies and/or television programs that Jennifer Aniston has appeared in. Narrative: Narrative: Relevant documents include movies and/or television programs that Jennifer Aniston has appeared in. </top>
Document Example <DOC><DOCNO>WTX010-B01-2</DOCNO><DOCOLDNO>IA B </DOCOLDNO><DOCHDR> text/html 264 HTTP/ OK Date: Sunday, 16-Feb-97 18:19:32 GMT Server: NCSA/SMI-1.0 MIME-version: 1.0 Content-type: text/html Last-modified: Friday, 02-Feb-96 19:51:15 GMT Content-length: 82 </DOCHDR> 1 Mr. Delleney did not participate in deliberation of this candidate. 1 Mr. Delleney did not participate in deliberation of this candidate.</DOC>
Link Information For approaches with PageRank/HITS For approaches with PageRank/HITS In-links In-links –“ A B C ” B and C contain links to A ex: WTX010-B WTX010-B WTX010-B Out-links Out-links –“ A B C ” A contains links pointed to B or C ex: WTX010-B WTX010-B01-89 WTX010-B01-119
Evaluation Evaluate top 100 retrieved documents Evaluate top 100 retrieved documents Evaluation metrics Evaluation metrics –Mean average precision (MAP) Use the program “ trec_eval” to evaluate system performance Use the program “ trec_eval” to evaluate system performance –Usage of trec_eval Usage of trec_evalUsage of trec_eval
Example Result for Evaluation (topic-num) (dummy) (docno) (rank) (score) (run-tag) 465Q0WTX017-B test 465 Q0WTX017-B test 465Q0WTX017-B test 465 Q0WTX017-B test 465 Q0WTX017-B test 465 Q0WTX018-B test 465 Q0WTX018-B test 465 Q0WTX012-B test 465 Q0WTX019-B test 465 Q0WTX019-B test 474 Q0WTX012-B test 474 Q0WTX017-B test 474 Q0WTX018-B test 474 Q0WTX013-B test 474 Q0WTX018-B test 474 Q0WTX015-B test 474 Q0WTX019-B test 474 Q0WTX014-B test 474 Q0WTX018-B test
Example of Relevance Judgments (topic-num) (dummy) (docno) (relevance) 4650WTX017-B WTX017-B WTX018-B WTX019-B WTX012-B WTX013-B WTX014-B WTX015-B WTX018-B WTX018-B
Summary of What to Do 1. Okapi-BM25 implementation (baseline) –With the fixed settings 2. Evaluate the baseline approach with training topics –using terms in as query 3. Survey or design your enhanced approach 4. Evaluate and optimize your approach with training topics 5. Submit report and demo with testing topics 6. Evaluate Okapi-BM25 and your approach with testing topics
Dataset Description (1/2) “ training_topics.txt” (file) “ training_topics.txt” (file) –30 topics for system development “ qrels_training_topics.txt” (file) “ qrels_training_topics.txt” (file) –Relevance judgments for training topics “ documents ” (directory) “ documents ” (directory) –Including 10.rar files of raw documents “ in_links.txt” (file) “ in_links.txt” (file) –In-link information “ out_links.txt ” (file) “ out_links.txt ” (file) –Out-link information
Dataset Description (2/2) “ trec_eval.exe ” (file) “ trec_eval.exe ” (file) –Binary evaluation program “ trec_eval.8.1.rar” (file) “ trec_eval.8.1.rar” (file) –Source of trec_eval for making in UNIX