Is Question Answering an Acquired Skill? Soumen Chakrabarti IIT Bombay With Ganesh Ramakrishnan Deepa Paranjpe Vijay Krishnan Arnab Nandi.

Slides:



Advertisements
Similar presentations
An Ontology Creation Methodology: A Phased Approach
Advertisements

ThemeInformation Extraction for World Wide Web PaperUnsupervised Learning of Soft Patterns for Generating Definitions from Online News Author Cui, H.,
Is Question Answering an Acquired Skill? Soumen Chakrabarti G. Ramakrishnan D. Paranjpe P. Bhattacharyya IIT Bombay.
Specialized models and ranking for coreference resolution Pascal Denis ALPAGE Project Team INRIA Rocquencourt F Le Chesnay, France Jason Baldridge.
Data Mining Feature Selection. Data reduction: Obtain a reduced representation of the data set that is much smaller in volume but yet produces the same.
RIPPER Fast Effective Rule Induction
Proceedings of the Conference on Intelligent Text Processing and Computational Linguistics (CICLing-2007) Learning for Semantic Parsing Advisor: Hsin-His.
Contextual Advertising by Combining Relevance with Click Feedback D. Chakrabarti D. Agarwal V. Josifovski.
Automatic indexing and retrieval of crime-scene photographs Katerina Pastra, Horacio Saggion, Yorick Wilks NLP group, University of Sheffield Scene of.
Using Web Queries for Learner Error Detection Michael Gamon, Microsoft Research Claudia Leacock, Butler-Hill Group.
Basic IR: Queries Query is statement of user’s information need. Index is designed to map queries to likely to be relevant documents. Query type, content,
Is Question Answering an Acquired Skill? Soumen Chakrabarti IIT Bombay With Ganesh Ramakrishnan Deepa Paranjpe Pushpak Bhattacharyya.
The Informative Role of WordNet in Open-Domain Question Answering Marius Paşca and Sanda M. Harabagiu (NAACL 2001) Presented by Shauna Eggers CS 620 February.
Using Information Extraction for Question Answering Done by Rani Qumsiyeh.
1 SIMS 290-2: Applied Natural Language Processing Marti Hearst Sept 22, 2004.
Learning syntactic patterns for automatic hypernym discovery Rion Snow, Daniel Jurafsky and Andrew Y. Ng Prepared by Ang Sun
A Framework for Named Entity Recognition in the Open Domain Richard Evans Research Group in Computational Linguistics University of Wolverhampton UK
Seven Lectures on Statistical Parsing Christopher Manning LSA Linguistic Institute 2007 LSA 354 Lecture 7.
Chapter 4 Query Languages.... Introduction Cover different kinds of queries posed to text retrieval systems Keyword-based query languages  include simple.
Information Retrieval in Practice
Empirical Methods in Information Extraction Claire Cardie Appeared in AI Magazine, 18:4, Summarized by Seong-Bae Park.
Probabilistic Model for Definitional Question Answering Kyoung-Soo Han, Young-In Song, and Hae-Chang Rim Korea University SIGIR 2006.
1 Statistical NLP: Lecture 10 Lexical Acquisition.
Query Processing: Query Formulation Ling573 NLP Systems and Applications April 14, 2011.
Processing of large document collections Part 2 (Text categorization) Helena Ahonen-Myka Spring 2006.
1 A study on automatically extracted keywords in text categorization Authors:Anette Hulth and Be´ata B. Megyesi From:ACL 2006 Reporter: 陳永祥 Date:2007/10/16.
Survey of Semantic Annotation Platforms
Interpreting Dictionary Definitions Dan Tecuci May 2002.
Text Classification, Active/Interactive learning.
AnswerBus Question Answering System Zhiping Zheng School of Information, University of Michigan HLT 2002.
Lecture 6 Hidden Markov Models Topics Smoothing again: Readings: Chapters January 16, 2013 CSCE 771 Natural Language Processing.
Annotating Words using WordNet Semantic Glosses Julian Szymański Department of Computer Systems Architecture, Faculty of Electronics, Telecommunications.
A Probabilistic Graphical Model for Joint Answer Ranking in Question Answering Jeongwoo Ko, Luo Si, Eric Nyberg (SIGIR ’ 07) Speaker: Cho, Chin Wei Advisor:
Word Sense Disambiguation in Queries Shaung Liu, Clement Yu, Weiyi Meng.
A Language Independent Method for Question Classification COLING 2004.
21/11/2002 The Integration of Lexical Knowledge and External Resources for QA Hui YANG, Tat-Seng Chua Pris, School of Computing.
Date: 2014/02/25 Author: Aliaksei Severyn, Massimo Nicosia, Aleessandro Moschitti Source: CIKM’13 Advisor: Jia-ling Koh Speaker: Chen-Yu Huang Building.
Reading Between The Lines: Object Localization Using Implicit Cues from Image Tags Sung Ju Hwang and Kristen Grauman University of Texas at Austin Jingnan.
Chapter 6: Information Retrieval and Web Search
Breaking through the syntax barrier: Searching with entities and relations Soumen Chakrabarti IIT Bombay
Collocations and Information Management Applications Gregor Erbach Saarland University Saarbrücken.
IBM Research © Copyright IBM Corporation 2005 | A Development Environment for Configurable Meta-Annotators in a Pipelined NLP Architecture Youssef Drissi,
Enhanced Answer Type Inference from Questions using Sequential Models Vijay Krishnan Sujatha Das Soumen Chakrabarti IIT Bombay.
Using a Named Entity Tagger to Generalise Surface Matching Text Patterns for Question Answering Mark A. Greenwood and Robert Gaizauskas Natural Language.
Deep Learning for Efficient Discriminative Parsing Niranjan Balasubramanian September 2 nd, 2015 Slides based on Ronan Collobert’s Paper and video from.
Number Sense Disambiguation Stuart Moore Supervised by: Anna Korhonen (Computer Lab)‏ Sabine Buchholz (Toshiba CRL)‏
Finding frequent and interesting triples in text Janez Brank, Dunja Mladenić, Marko Grobelnik Jožef Stefan Institute, Ljubljana, Slovenia.
Supertagging CMSC Natural Language Processing January 31, 2006.
Mining Dependency Relations for Query Expansion in Passage Retrieval Renxu Sun, Chai-Huat Ong, Tat-Seng Chua National University of Singapore SIGIR2006.
Language Modeling Putting a curve to the bag of words Courtesy of Chris Jordan.
4. Relationship Extraction Part 4 of Information Extraction Sunita Sarawagi 9/7/2012CS 652, Peter Lindes1.
Answer Mining by Combining Extraction Techniques with Abductive Reasoning Sanda Harabagiu, Dan Moldovan, Christine Clark, Mitchell Bowden, Jown Williams.
COT6930 Course Project. Outline Gene Selection Sequence Alignment.
Acquisition of Categorized Named Entities for Web Search Marius Pasca Google Inc. from Conference on Information and Knowledge Management (CIKM) ’04.
Using Wikipedia for Hierarchical Finer Categorization of Named Entities Aasish Pappu Language Technologies Institute Carnegie Mellon University PACLIC.
SEMANTIC VERIFICATION IN AN ONLINE FACT SEEKING ENVIRONMENT DMITRI ROUSSINOV, OZGUR TURETKEN Speaker: Li, HueiJyun Advisor: Koh, JiaLing Date: 2008/5/1.
Identifying “Best Bet” Web Search Results by Mining Past User Behavior Author: Eugene Agichtein, Zijian Zheng (Microsoft Research) Source: KDD2006 Reporter:
1 Question Answering and Logistics. 2 Class Logistics  Comments on proposals will be returned next week and may be available as early as Monday  Look.
Learning Analogies and Semantic Relations Nov William Cohen.
Learning Event Durations from Event Descriptions Feng Pan, Rutu Mulkar, Jerry R. Hobbs University of Southern California ACL ’ 06.
Feature Assignment LBSC 878 February 22, 1999 Douglas W. Oard and Dagobert Soergel.
Selecting Relevant Documents Assume: –we already have a corpus of documents defined. –goal is to return a subset of those documents. –Individual documents.
Integrating linguistic knowledge in passage retrieval for question answering J¨org Tiedemann Alfa Informatica, University of Groningen HLT/EMNLP 2005.
Question Classification II
Text Based Information Retrieval
Introduction to Information Retrieval
Template-based Question Answering over RDF Data
Johns Hopkins 2003 Summer Workshop on Syntax and Statistical Machine Translation Chapters 5-8 Ethan Phelps-Goodman.
Memory-Based Learning Instance-Based Learning K-Nearest Neighbor
Presentation transcript:

Is Question Answering an Acquired Skill? Soumen Chakrabarti IIT Bombay With Ganesh Ramakrishnan Deepa Paranjpe Vijay Krishnan Arnab Nandi

QAChakrabarti Web search and QA  Information need – words relating “things” + “thing” aliases = telegraphic Web queries Cheapest laptop with wireless  best price laptop Why is the sky blue?  sky blue reason When was the Space Needle built?  “Space Needle” history  Entity + relation extraction technology better than ever (SemTag, KnowItAll, Biotext) Ontology extension (e.g., is a kind of) List extraction (e.g., is an instance of) Slot-filling (author X wrote book Y)

QAChakrabarti Factoid QA  Specialize given domain to a token related to ground constants in the query What animal is Winnie the Pooh? hyponym(“animal”) NEAR “Winnie the Pooh” When was television invented? instance-of(“time”) NEAR “television” NEAR synonym(“invented”)  FIND x “NEAR” GroundConstants(question) WHERE x IS-A Atype(question) Ground constants: Winnie the Pooh, television Atypes: animal, time

QAChakrabarti A relational view of QA  Entity class or atype may be expressed by A finite IS-A hierarchy (e.g. WordNet, TAP) A surface pattern matching infinitely many strings (e.g. “digit+”, “Xx+”, “preceded by a preposition”)  Match selectors, specialize atype to answer tokens QuestionAtype clues Selectors Answer passage Question words “Answer zone” Direct syntactic match Entity class IS-A Limit search to certain rows Locate which column to read “Answer zone” Attribute or column name

QAChakrabarti Benefits of the relational view  “Scaling up by dumbing down” Next stop after vector-space Far short of real knowledge representation and inference Barely getting practical at (near) Web scale  Can set up as a learning problem: train with questions (query logs) and answers in context  Transparent, self-tuning, easy to deploy Feature extractors used in entity taggers Relational/graphical learning on features

QAChakrabarti What TREC QA feels like  How to assemble chunker, parser, POS and NE tagger, WordNet, WSD, … into a QA system?  Experts get much insight from old QA pairs Matching an upper-cased term adds a 60% bonus … for multi-words terms and 30% for single words Matching a WordNet synonym … discounts by 10% (lower case) and 50% (upper case) Lower-case term matches after Porter stemming are discounted 30%; upper-case matches 70%

QAChakrabarti Talk outline  Relational interpretation of QA  Motivation for a “clean-room” IE+ML system  Learning to map between questions and answers using is-a hierarchies and IE-style surface patterns Can handle prominent finite set of atypes: person, place, time, measurements,…  Extending to arbitrary atype specializations Required for what… and which… questions  Ongoing work and concluding remarks

QAChakrabarti Feature + Soft match  FIND x “NEAR” GroundConstants(question) WHERE x IS-A Atype(question)  No fixed question or answer type system  Convert “x IS-A Atype(question)” to a soft match “DoesAtypeMatch(x, question) QuestionAnswer tokens Passage IE-style surface feature extractors WordNet hypernym feature extractors IE-style surface feature extractors Question feature vector Snippet feature vector Learn joint distrib.

QAChakrabarti Feature extraction: Intuition howwho fastmanyfarrich wrotefirst How fast can a cheetah run? A cheetah can chase its prey at up to 90 km/h How fast does light travel? Nothing moves faster than 186,000 miles per hour, the speed of light rate#n#2 abstraction#n#6 NNS rate#n#2 magnitude_relation#n#1 mile#n#3 linear_unit#n#1 measure#n#3 definite_quantity#n#1 paper_money#n#1 currency#n#1 writer, composer, artist, musician NNP, person explorer

QAChakrabarti Feature extractors  Question features: 1, 2, 3-token sequences starting with standard wh-words  Passage surface features: hasCap, hasXx, isAbbrev, hasDigit, isAllDigit, lpos, rpos,…  Passage WordNet features: all noun hypernym ancestors of all senses of token  Get top 300 passages from IR engine  For each token invoke feature extractors  Label = 1 if token is in answer span, 0 o/w  Question vector x q, passage vector x p

QAChakrabarti Preliminary likelihood ratio tests Surface patternsWordNet hypernyms

QAChakrabarti A simple, flat conditional model  Let x = x q  x p (pairwise product of elems)  Model Pr(Y=1|x) = exp(w  x)/(1+exp(w  x))  For every question-feature, passage-feature pair, w has a parameter  Expect to perform better than “linear” model x=(x p,x q )  Can discount for redundancy in pair info  If x q (x p ) is fixed, what x p (x q ) will yield the largest Pr(Y=1|x)? (linear iceberg query) how_far when what_city region#n#3 entity#n#1

QAChakrabarti Classification accuracy  Pairing more accurate than linear model  Steep learning curve; linear never “gets it” beyond “prior” atypes like proper nouns (common in TREC)  Are the estimated w parameters meaningful?

QAChakrabarti Parameter anecdotes  Surface and WordNet features complement each other  General concepts get negative params: use in predictive annotation  Learning is symmetric (Q  A)

QAChakrabarti Query-driven information extraction  “Basis” of atypes A, a  A could be a synset, a surface pattern, feature of a parse tree  Question q “projected” to vector (w a : a  A) in atype space via learning conditional model  E.g. if q is “when…” or “how long…” w hasDigit and w time_period#n#1 are large, w region#n#1 is small  Each corpus token t has associated indicator features  a (t ) for every a  E.g.  hasDigit (3,000) =  is-a(region#n#1) (Japan) = 1  Can also learn [0,1] value of is-a proximity

QAChakrabarti Single token scoring  A token t is a candidate answer if  H q (t ): Reward tokens appearing “near” selectors matched from question 0/1: appears within fixed window with selector/s Activation in linear token sequence model Proximity in chunk sequences, parse trees,…  Order tokens by decreasing Atype indicator features of the token Projection of question to “atype space” …the armadillo, found in Texas, is covered with strong horny plates

QAChakrabarti Mean reciprocal rank (MRR)  n q = smallest rank among answer passages  MRR = (1/|Q|)  q  Q (1/n q ) Dropping passage from #1 to #2 as bad as dropping it from #2 to   TREC requires MRR5: round up n q >5 to  Improving rank from 20 to 6 as useless as improving it from 20 to 15  Aggregate score influenced by many complex subsystems Complete description rarely available

QAChakrabarti Effect of eliminating non-answers  300 top IR score hits  If Pr(Y=1|token) < threshold reject token  All tokens rejected then reject passage  Present survivors in IR order

QAChakrabarti Drill-down and ablation studies  Scale average MRR improvement to 1 What, Which < average Who  average  Atype of what… and which… not captured well by 3-grams starting at wh-words  Atype ranges over essentially infinite set with relatively little training data

QAChakrabarti Talk outline  Relational interpretation of QA  Motivation for a “clean-room” IE+ML system  Learning to map between questions and answers using is-a hierarchies and IE-style surface patterns Can handle prominent finite set of atypes: person, place, time, measurements,…  Extending to arbitrary atype specializations Required for what… and which… questions  Ongoing work and concluding remarks

QAChakrabarti What…, which…, name… atype clues  Assumption: Question sentence has a wh- word and a main/auxiliary verb  Observation: Atype clues are embedded in a noun phrase (NP) adjoining the main or auxiliary verb  Heuristic: Atype clue = head of this NP Use a shallow parser and apply rule  Head can have attributes Which (American (general)) is buried in Salzburg? Name (Saturn’s (largest (moon)))

QAChakrabarti Atype clue extraction stats  Simple heuristic quite effective  If successful, extracted atype is mapped to WordNet synset (moon  celestial body etc.)  If no atype of this form available, try the “self- evident” atypes (who, when, where, how_X etc.)  New boolean feature for candidate token: is token hyponym of atype synset?

QAChakrabarti The last piece: Learning selectors  Which question words are likely to appear (almost) unchanged in an answer passage? Constants in select-clauses of SQL queries Guides backoff policy for keyword query  Local and global features POS of word, POS of adjacent words, case info, proximity to wh-word Suppose word is associated with synset set S NumSense: size of S (how polysemous is the word?) NumLemma: average #lemmas describing s  S

QAChakrabarti Selector results  Global features (IDF, NumSense, NumLemma) essential for accuracy Best F1 accuracy with local features alone: 71—73% With local and global features: 81%  Decision trees better than logistic regression F1=81% as against LR F1=75% Intuitive decision branches But logistic regression gives scores for query backoff

QAChakrabarti Putting together a QA system QA System Wordnet POS Tagger Training Corpus Shallow parser Learning tools N-E Tagger

QAChakrabarti Question Passage Index Corpus Sentence splitter Passage indexer Candidate passage Keyword query Keyword query generator Shallow Parser Noun and verb markers Atype Extractor Atype clues Learning to rerank passages Sample features: Do selectors match? How many? Is some non-selector passage token a specialization of the question’s atype clue? Min, avg, linear token distance between candidate token and matched selectors Learning to rerank passages Sample features: Do selectors match? How many? Is some non-selector passage token a specialization of the question’s atype clue? Min, avg, linear token distance between candidate token and matched selectors Logistic Regression Reranked passages Putting together a QA system Tokenizer POS Tagger Tagged question Tokenizer POS Tagger Entity Extractor Tagged passage Selector Learner Is QA pair?

QAChakrabarti Learning to re-rank passages  Remove passage tokens matching selectors User already knows these are in passage  Find passage token/s specializing atype  For each candidate token collect Atype of question, original rank of passage Min, avg linear distances to matched selectors POS and entity tag of token if available Ushuaia, a port of about 30,000 dwellers set between the Beagle Channel and … How many inhabitants live in the town of Ushuaia selector match Surface pattern hasDigits WordNet match 5 tokens apart1

QAChakrabarti Re-ranking results  Categorical and numeric attributes  Logistic regression  Good precision, poor recall  Use logit score to re-rank passages  Rank of first correct passage shifts substantially

QAChakrabarti MRR gains from what, which, name  Substantial gain in MRR  What/which now show above-average MRR gains  TREC 2000 top MRRs:

QAChakrabarti Generalization across corpora  Across-year numbers close to train/test split on a single year  Features and model seem to capture corpus- independent linguistic Q+A artifacts

QAChakrabarti Conclusion  Clean-room QA= feature extraction+learning Recover structure info from question Learn correlations between question structure and passage features  Competitive accuracy with negligible domain expertise or manual intervention  Ongoing work Model how selector and atype are related Model coefficients to predictive annotation Combine token scores to better passage scores Treat all question types uniformly Use redundancy available from the Web