Published on November 16, 2007
Question Answering: Question Answering Zhikun Meng Overview: Overview What is Question Answering? Why Question Answering? How Question Answering? What is the current status of Question Answering? How to evaluate Question Answering? What is the future of Question Answering? What is Question Answering?: What is Question Answering? What is question answering system? Question answer clear What is Question Answering?: What is Question Answering? Answer: Question answering systems are designed to find answers to open domain questions in a large collection of documents. Why Question Answering?: Why Question Answering? How Question Answering?: How Question Answering? General architecture question Question Classification Information Retrieval Answer Extraction Answer answer e.g. What is Calvados? /Q is /A where:/Q=“(Calvados)” Query=“Calvados is” Text retrieva l=“…Calvados is often used in cooking…Calvados is a dry apple brandy made in… /A is : a dry apple brandy Answer: /Q is /A: “Calvados” is ”a dry apple brandy” Question Classification: Question Classification e.g. “How much could you rent a Volkswagen bug for in 1966?” Key word preprocessing (split/spell check/normalize) Volkswagen-Volkswagen; “Rotary engine cars were made by what company?” - “What company were rotary engine cars made by?” Question Classification: Question Classification 2. Construction of question representation How much: Question stem rent: Answer type term 1966:Data constraint Volkswagen bug Question Classification: Question Classification 3.Derivation of answer type “How much”+ “rent” ->Money Question Classification: Question Classification 4.Key word selection Volkswagen AND bug AND rent Question Classification: Question Classification 5.Key word expansion rent-rented Information Retrieval: Information Retrieval Retrieval documents and passages: query: Volkswagen AND bug AND rent The retrieval engine returns the documents containing all keywords (e.g.60 document passages from 1,000,000 documents collection) Information Retrieval: Information Retrieval 2. Passage filtering date constraint 1966. Out of the 60 passages returned by the retrieval engine for Q013, two passages are retained after passage post filtering. Answer Extraction: Answer Extraction Identification of candidate answers Answer type: Money Identified candidates include $1 and USD 520. Answer Extraction: Answer Extraction 2. Answer Ranking score: $1 USD 520 Answer Extraction: Answer Extraction 3. Answer formulation rent a Volkswagen bug for $1 a day What is the current status of Question Answering?: What is the current status of Question Answering? Text REtrieval Conference (TREC) http://trec.nist.gov/ Cross Language Evaluation Forum (CLEF) http://clef.isti.cnr.it/ NII-NACSIS Test Collection for IR Systems Project (NTCIR) http://research.nii.ac.jp/ntcir/index-en.html TREC-8: TREC-8 In 1999, the 8th Text REtrieval Conference (TREC-8) first proposed QA track . TREC-8 tasks included: Answer factoid questions by returning a text snippet which contained an answer to the question Build a reusable QA test collection. TREC-9: TREC-9 Comparing to TREC-8, in TREC-9 the biggest change was the switch to “real" questions, rather than questions created especially for the track. The absolute value of scores dropped, yet the performance of the TREC-9 systems improved significantly in QA technology. TREC 2001 : TREC 2001 The major adjustment of TREC 2001 QA track is to divide the task into three separate tasks: the main task, the list task the context task. TREC 2002 : TREC 2002 TREC 2002 QA track contained two tasks, the main task and the list task The system was required to return the exact answers Systems were limited to one response per question, not five Ranking metric changed to the confidence-weighted score TREC 2003 : TREC 2003 The TREC 2003 question answering track contained two tasks: the passages task (factoid) the main task (factoids, lists, definitions) significant participation was involved in lists question task TREC 2004: TREC 2004 The factoid question and list question are not independent; instead, they are all related to given topics More participants are involved in resolving list question and definition question tasks. CLEF: CLEF In 2002, CLEF proposed its own Question Answering track which focuses on European Language QA, especially Cross-lingual QA (CLQA). CLEF 2003 QA track was divided into monolingual and bilingual tasks NTCIR: NTCIR In 2003, the Asian information retrieval conference NTCIR proposed its Question Answering track which focuses on Asian Language QA. How to evaluate Question Answering?: How to evaluate Question Answering? TREC About 1500 question collections from TREC-8,9,2001 Answers were extracted from a 3-Gbyte text collection containing about 1 million documents from sources such as the Los Angeles Times and Wall Street Journal. Each answer has at most 50 characters. The answer accuracy used to be measured by the Mean Reciprocal Rank (MRR) metric used by NIST in the TREC QA evaluations. In TREC2002, ranking metric was changed to the confidence-weighted score. How to evaluate Question Answering?: How to evaluate Question Answering? CLEF In CLEF 2002, a test set for future cross-lingual research, DISEQuA (Dutch Italian Spanish English Questions and Answers )Corpus, is created. What is the future of Question Answering?: What is the future of Question Answering? Factoid Questions Lists Questions Definition Questions Non-English monolingual QA Cross-lingual QA cross-lingual and cross-media QA Summary: Summary QA definition General QA system architecture QA technologies QA research milestone Roadmap Questions?: Questions?