Question Answering Passage Retrieval Using Dependency Parsing Hang Cui Renxu Sun Keya Li Min-Yen Kan Tat-Seng Chua Department of Computer Science National University of Singapore August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 1/28 Passage Retrieval in Question Answering QA System Document Retrieval • To narrow down the search scope • Can answer questions with more context Passage Retrieval Answer Extraction August 17, 2005 • Lexical density based • Distance between question words Question Answering Passage Retrieval Using Dependency Parsing 2/28 Density Based Passage Retrieval Method • However, density based can err when … <Question> What percent of the nation's cheese does Wisconsin produce? Incorrect: … the number of consumers who mention California when asked Relationships between about cheese has risen by 14 percent, while the number specifying Wisconsin has dropped 16 percent. matched words differ … Incorrect: The wry “It's the Cheese” ads, which attribute California's allure to its cheese _ and indulge in an occasional dig at the Wisconsin stuff'' … sales of cheese in California grew three times as fast as sales in the nation as a whole 3.7 percent compared to 1.2 percent, … Incorrect: Awareness of the Real California Cheese logo, which appears on about 95 percent of California cheeses, has also made strides. Correct: In Wisconsin, where farmers produce roughly 28 percent of the nation's cheese, the outrage is palpable. August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 3/28 Our Solution • Examine the relationship between words – Dependency relations • Exact match of relations for answer extraction • Has low recall because same relations are often phrased differently • Fuzzy match of dependency relationship – Statistical similarity of relations August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 4/28 Measuring Sentence Similarity Sim (Sent1, Sent2) = ? Sentence 1 Sentence 2 Matched words Lexical matching + Similarity of relations between matched words Similarity of individual relations August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 5/28 Outline • • • • • Extracting and Paring Relation Paths Measuring Path Match Scores Learning Relation Mapping Scores Evaluations Conclusions August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 6/28 Outline • • • • • Extracting and Paring Relation Paths Measuring Path Match Scores Learning Relation Mapping Scores Evaluations Conclusions August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 7/28 What Dependency Parsing is Like • Minipar (Lin, 1998) for dependency parsing • Dependency tree Root ad wh n he – Nodes: words/chunks in the sentence – Edges (ignoring the direction): labeled by relation types produce p re de t percent subj p what of Wisconsin p co mpn What percent of the nation's cheese does Wisconsin produce? cheese ge n nation August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 8/28 Extracting Relation Paths • Relation path – Vector of relations between two nodes in the tree Root ad wh n he produce p re de t percent subj p what of produce < P1: subj > Wisconsin percent < P2: prep pcomp-n > cheese Wisconsin p co Two constraints for relation paths: 1. Path length (less than 7 relations) 2. Ignore those between two words that are within a chunk, e.g. New York. mpn cheese ge n nation August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 9/28 Paired Paths from Question and Answer In Wisconsin, where farmers produce roughly 28 percent of the nation's cheese, the outrage is palpable. What percent of the nation's cheese does Wisconsin produce? Root mo n ad wh i he d Root produce in produce percent subj p mpn < P1(Sent) : pcomp-n mod i > farmers Wisconsin Wisconsin mo d of 28 percent p co t de p re what su b j obj < P1(Q) : subj > of p co n p co m p mp- -n Paired Relation Paths cheese cheese n nation nation August 17, 2005 gen ge SimRel (Q, Sent) = ∑i,j Sim (Pi (Q), Pj(Sent)) Question Answering Passage Retrieval Using Dependency Parsing 10/28 Outline • • • • • Extracting and Paring Relation Paths Measuring Path Match Scores Learning Relation Mapping Scores Evaluations Conclusions August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 11/28 Measuring Path Match Degree • Employ a variation of IBM Translation Model 1 – Path match degree (similarity) as translation probability • MatchScore (PQ, PS) → Prob (PS | PQ ) • Relations as words • Why IBM Model 1? – No “word order” – bag of undirected relations – No need to estimate “target sentence length” • Relation paths are determined by the parsing tree August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 12/28 Calculating Translation Probability (Similarity) of Paths Given two relation paths from the question and a candidate sentence Prob( PS | PQ ) mn m m n (S ) (Q ) Pt ( Reli | Reli ) 11 n 1 i 1 Considering the most probable alignment (finding the most probable mapped relations) Prob( PS | PQ ) m n (S ) (Q ) P ( Re l | Re l ) t i A n i i 1 Take logarithm and ignore the constants (for all sentences, question path length is a constant) MatchScore ( PS ) Prob( PS | PQ ) August 17, 2005 ' n n log Pt ( Reli( S ) | Rel A(Qi ) ) i 1 MatchScores of paths are combined to give the sentence’s relevance to the question. ? Question Answering Passage Retrieval Using Dependency Parsing 13/28 Outline • • • • • Extracting and Paring Relation Paths Measuring Path Match Scores Learning Relation Mapping Scores Evaluations Conclusions August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 14/28 Training and Testing Testing Training Sim ( Q, Sent ) = ? Prob ( PSent | PQ ) = ? 1. Mutual Q - A pairs Similarityinformation between (MI) relationbased vectors 2. Expectation Maximization (EM)between based Paired Relation Paths Similarity individual relations P ( Rel (Sent) | Rel (Q) ) = ? Relation Mapping Model Relation Mapping Scores August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 15/28 Approach 1: MI Based • Measures bipartite co-occurrences in training path pairs • Accounts for path length (penalize those long paths) • Uses frequencies to approximate mutual information Pt ( MI ) ( Reli( S ) | Rel (j Q ) ) log August 17, 2005 (Q ) (S ) ( Rel j , Reli ) | Rel (j Q ) | | Reli( S ) | Question Answering Passage Retrieval Using Dependency Parsing 16/28 Approach – 2: EM Based • Employ the training method from IBM Model 1 – Relation mapping scores = word translation probability – Utilize GIZA to accomplish training – Iteratively boosting the precision of relation translation probability • Initialization – assign 1 to identical relations and a small constant otherwise August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 17/28 Outline • • • • Extracting and Paring Relation Paths Measuring Path Match Scores Learning Relation Mapping Scores Evaluations – – – – Can relation matching help? Can fuzzy match perform better than exact match? Can long questions benefit more? Can relation matching work on top of query expansion? • Conclusions August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 18/28 Evaluation Setup • Training data – 3k corresponding path pairs from 10k QA pairs (TREC-8, 9) • Test data – 324 factoid questions from TREC-12 QA task • Passage retrieval on top 200 relevant documents by TREC August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 19/28 Comparison Systems • MITRE –baseline – Stemmed word overlapping – Baseline in previous work on passage retrieval evaluation • SiteQ – top performing density based method – using 3 sentence window • NUS – Similar to SiteQ, but using sentences as passages • Strict Matching of Relations – Simulate strict matching in previous work for answer selection – Counting the number of exactly matched paths • Relation matching are applied on top of MITRE and NUS August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 20/28 Evaluation Metrics • Mean reciprocal rank (MRR) – On the top 20 returned passages – Measure the mean rank position of the correct answer in the returned rank list • Percentage of questions with incorrect answers • Precision at the top one passage August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 21/28 Performance Evaluation • All improvements are statistically significant (p<0.001) MRR Comparison • MI and EM do not make much difference given our training data el at io St n( ri c M IT t_ R R el E ) at i on R el (N at U io S) n_ R el M at I (M io n_ IT SM RE T1 ) (M R el IT at R io E) n_ R el M at I( io NU n_ S SM ) T1 (N U S) S Rel_EM (NUS) 0.3625 0.4161 0.4218 0.4756 0.4761 N/A N/A N/A +38.26 N/A N/A +33.88 N/A N/A +49.50 +8.14 +11.69 +81.25 +31.10 +35.41 +108.09 +50.50 +55.43 +110.94 +52.57 +57.56 +137.85 +72.03 +77.66 +138.08 +72.19 +77.83 % Incorrect 45.68% 37.65% 33.02% 41.96% 32.41% 29.63% 29.32% 24.69% 24.07% Precision at top one passage 0.1235 0.1975 0.1759 0.2253 0.2716 0.3364 0.3457 0.3889 0.3889 August 17, 2005 St ri c t_ R N U 0.2990 Si te Q 0.2677 % MRR improvement over MITRE SiteQ NUS 0.2000 Rel_MI (NUS) 0.2765 M IT R MRR E MRR 0.60 – EM needs more training data Fuzzy matching 0.50 – MI is more susceptible to noise, so may not scale well outperforms strict 0.40 0.30 matching significantly. Passage 0.20 retrieval Rel_Strict Rel_Strict Rel_MI Rel_EM MITRE SiteQ NUS systems 0.10 (MITRE) (NUS) (MITRE) (MITRE) 0.00 Question Answering Passage Retrieval Using Dependency Parsing 22/28 Performance Variation to Question Length • Long questions, with more paired paths, tend to improve more – Using the number of non-trivial question terms to approximate question length 0.65 0.6 0.55 MRR 0.5 Rel_NUS_EM 0.45 0.4 Rel_MITRE_EM 0.35 0.3 0.25 0.2 0 2 4 6 8 # Question Term s August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 23/28 Error Analysis • Mismatch of question terms • e.g. In which city is the River Seine • Introduce question analysis • Paraphrasing between the question and the answer sentence • e.g. write the book → be the author of the book • Most of current techniques fail to handle it • Finding paraphrasing via dependency parsing (Lin and Pantel) August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 24/28 Performance on Top of Query Expansion • On top of query expansion, fuzzy relation matching brings a further 50% improvement • However – query expansion doesn’t help much on a fuzzy relation matching system – Expansion terms do not help in paring relation paths Passage Retrieval Systems NUS (baseline) NUS+QE Rel_MI (NUS+QE) Rel_EM (NUS+QE) MRR (% improvement over baseline) 0.2677 0.3293 (+23.00%) 0.4924 (+83.94%) 0.4935 (+84.35%) % MRR improvement over NUS+QE N/A N/A +49.54% +49.86% % Incorrect 33.02% 28.40% 22.22% 22.22% Precision at top one passage 0.1759 0.2315 0.4074 0.4074 August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing Rel_EM (NUS) 0.4761 25/28 Outline • • • • • Extracting and Paring Relation Paths Measuring Path Match Scores Learning Relation Mapping Scores Evaluations Conclusions August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 26/28 Conclusions • Proposed a novel fuzzy relation matching method for factoid QA passage retrieval – Brings dramatic 70%+ improvement over the state-ofthe-art systems – Brings further 50% improvement over query expansion – Future QA systems should bring in relations between words for better performance • Query expansion should be integrated to relation matching seamlessly August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 27/28 Q&A Thanks! August 17, 2005 Question Answering Passage Retrieval Using Dependency Parsing 28/28