Presentation is loading. Please wait.

Presentation is loading. Please wait.

Relevance Models for QA Project Update University of Massachusetts, Amherst AQUAINT meeting December, 2002 Bruce Croft and James Allan, PIs.

Similar presentations


Presentation on theme: "Relevance Models for QA Project Update University of Massachusetts, Amherst AQUAINT meeting December, 2002 Bruce Croft and James Allan, PIs."— Presentation transcript:

1 Relevance Models for QA Project Update University of Massachusetts, Amherst AQUAINT meeting December, 2002 Bruce Croft and James Allan, PIs

2 UMass AQUAINT Project Status Question answering using language models  Carried out more experiments using basic LM approach  Developed new model(s) and starting more experiments  Moved experiments to LEMUR toolkit Query triage  Studied Clarity measure for questions Question answering with semi-structured data  Developed HMM and CRF-based table extractors  More experiments on question answering with table structure Answer updating  Experiments with time-based questions

3 QA using LM P(Answer|Question) can be estimated many ways  Could be done directly, but usually will involve intermediate steps such as documents, question classes  Initially focused on answer passages, but “extracted” answers can be modeled  Can model “templates” as well as n-gram answer models  Can also introduce cross-lingual QA through P(A lang1 |Q lang2 ) Every approach requires training data  “answer mining” for answer models/templates  incorporating user feedback

4 Query Triage Given a question, what can we infer from it?  Query vs. question  Quality (does it need to be made more precise)  Type (likely form of answers and granularity)  Human intermediation (should it be directed to a human expert?) Previous work developed “Clarity” measure for queries and tested on TREC ad-hoc data  Demonstrated high correlation with performance  Threshold can be set automatically Current research focuses on TREC QA data

5 Basic result: We can predict question performance (with some qualifications)  Did not work for some TREC question classes For example:  What is the date of Bastille Day? TREC-9P Clarity score 2.49  What time of year do most people fly? TREC-9P Clarity score 0.76 Predicting Question Performance

6 “the” “do”, “day”, “what” “celebrate” “paris” “bastille” “assmann” terms Log P Clarity score computation Question Q, text Question Q, text... Passages, A... Passages ranked by P(A|Q) retrieve model passage collection language model passage collection language model question- related language model question- related language Compute divergence Clarity Score

7 Clarity Example (for queries) term rank p q Log 2 (p q /p c ) Top 6 terms in query model: 1. "adjust" 2. "federal" 3. "action" 4. "land" 5. "occur" 6. "hyundai" 56.08 "What adjustments should be made when federal action occurs?" (clar. 0.37) 56.12 "Show me predictions for changes in the prime lending rate and any changes made in the prime lending rates" (clar. 2.85) Top 6 terms in query model: 1. "bank" 2. "hong" 3. "kong" 4. "rate" 5. "lend" 6. "prime"

8 Test System Passages:  Two sentences, overlapping  from top retrieved docs for all questions Measuring performance:  Question Likelihood used to rank passages  Average precision (rather than MRR)  Top 8 documents to estimate Clarity scores

9 Precision vs. Clarity (Time Qs) Average Precision What is the date of Bastille Day? What time of year do most people fly? What is Martin Luther King Jr 's real birthday? Clarity Score

10 Question Type# of QsRank Correlation (R) P-Value Amount 33– 0.132 0.77 Famous 740.197 0.046 Location 1000.386 0.000062 Person 93– 0.109 0.85 Time 470.458 0.00094 Miscellaneous 1300.355 0.000028 Correlation by Question Type

11 Strong on average Allows prediction of question performance Variation with question type  Two bad (R<0) cases: Amount and Person Amount: only has 33 questions, only a few bad Qs Person: 93 questions, plenty of bad Qs to analyze What’s going on? Correlation Analysis

12 Two kinds of mistakes: High clarity, low average precision  E.g. What is Martin Luther King Jr 's real birthday?  Answerless, coherent, very likely context in collection  Rare (good thing for the method) Low clarity, high average precision  Various kinds of bad luck  Often coupled with few relevant passages  Many examples in Person case… Predictive Mistakes 0 3 1 0 Ave. Precision Clarity Score

13 Precision vs. Clarity (Person Qs) 15 “really bad” mistakes  “Really bad” ≡ clarity score 70 %-ile  8 with many relevant answer passages ( > 50 ) 5 (one-third) are slight variants of Who created “The Muppets”? 2 variants of What king signed the Magna Carta? 1 other question with plenty of relevants  7 with few relevant answer passages E.g. Silly Putty was invented by whom?, 2 rels 0 0.2 0.4 0.6 0.8 1 1.21.622.42.8 Ave. Precision Clarity Score

14 QA using Tables Developed and tested QUASM demonstration system using non-LM techniquesQUASM demonstration system  extraction of tabular structure  answer passages constructed from extracted data and metadata  extension of question types for “statistical” data  failure analysis Major focus now is to develop probabilistic framework for whole process  tabular structure extraction  answer passage representation  P(Answer|Question)

15 QuASM – Lessons Learned Much harder to find answers in tables than in text Table extraction is the key issue Representation of answer passages also very important  what is an answer passage for tables?  e.g. too much metadata can cause poor retrieval

16 Table Extraction Heuristics do a good job of identifying tables  97.8% percent of lines labeled correctly as in or out of table Small labeling errors, however, can lead to poor retrieval Current algorithm for extracting header information too permissive

17 Text Table Transformation Number and Percent of Children under 19 Years of Age, at or below 200 Percent of Poverty, by State: Three-Year Averages for 1997, 1998, and 1999. (Numbers in Thousands) _________________________________________________________________________________ | AT OR BELOW | AT OR BELOW 200% OF POVERTY | Total children | 200% OF POVERTY | WITHOUT HEALTH INSURANCE | under 19 years, |____________________________|_____________________________| all income levels | Standard Standard| Standard Standard | |Number error Pct. error |Number error Pct. error | ______________________|____________________________|_____________________________| Alabama....... 1,114 | 499 45.8 44.6 3.1 | 106 21.3 9.6 1.8 | Alaska........ 215 | 63 6.4 29.4 2.5 | 18 3.4 8.3 1.5 | Arizona....... 1,430 | 730 54.7 51.1 2.7 | 272 33.6 19.0 2.1 | Arkansas...... 740 | 377 30.5 50.5 2.9 | 111 16.5 14.7 2.0 |

18 Text Table Transformation - Problems (Numbers in Thousands) | AT OR BELOW | AT OR BELOW 200% OF POVERTY | Total children | 200% OF POVERTY | WITHOUT HEALTH INSURANCE | under 19 years, |____________________________|_____________________________| all income levels | Standard Standard| Standard Standard | |Number error Pct. error |Number error Pct. error | Alabama....... AT OR BELOW 200% OF POVERTY ____________________________ Standard Number. | 499 Missed part of title due to lack of indentation Extraneous text

19 New Labeling 3 Cells 2 Gaps Mostly Letters Mostly Digits Header Like Dashes Starts with Spaces Consecutive Spaces All White Space Features NONTABLE BLANKLINE TITLE SUPERHEADER TABLEHEADER SUBHEADER DATAROW SEPARATOR SECTIONHEADER SECTIONDATAROW TABLEFOOTNOTE TABLECAPTION Line Tags

20 Text Table Extraction Model Non-Table TitleData Row Super Header Table Header Subheader Finite State Machine (hidden Markov process) Non-TableTitleSuper HeaderTable HeaderData Row Visible feature vectors probabilistically infer state sequence.

21 Features for Table Extraction These features are not independent  Many correlations  Overlapping and long- distance dependencies  Observations from the past and future 3 Cells 2 Gaps Mostly Letters Mostly Digits Header Like Dashes Starts with Spaces Consecutive Spaces All White Space Features

22 Non-TableTitleSuper HeaderTable HeaderData Row Observations are conditioned on state  HMMs are the standard sequence model  They are a generative model of the sequence  Generative models do not easily handle non-independent features. Hidden Markov Models

23 Conditional Random Fields Non-TableTitleSuper HeaderTable HeaderData Row State sequence is conditioned on entire observation sequence. A conditional model:  Can examine features, but is not responsible for generating them.  Doesn’t have to explicitly model their dependencies.  Has the ability to handle many arbitrary features with the full power of finite state automata.

24 Results ExperimentPercentage of Lines Labeled Correctly Random, Training Data MLE11.4% HMM83.0% Fully Connected CRF93.3% Original Heuristic (4 labels)77.0% Label six test documents, total of 5817 lines.

25 Summary of Plans Testing a probabilistic model for QA Refining the Clarity measure for questions Finer-grain table extraction and QA tests Time-dependent language models


Download ppt "Relevance Models for QA Project Update University of Massachusetts, Amherst AQUAINT meeting December, 2002 Bruce Croft and James Allan, PIs."

Similar presentations


Ads by Google