1
University of Texas at Dallas
Computational Implicatures for Advanced Question Answering
Sanda Harabagiu, Alessandro Moschitti,
Adrian Atanasiu, Paul Morarescu
2
University of Texas at Dallas
Question Processing There are many reasons for which current
QA systems cannot accurately produce answers:
1. Questions are too complex – need translations in sets of simpler questions
2. Sometimes implicit knowledge is presupposed- Pragmatic knowledge of the domain - Recognition of non-literal expressions, followed by
coercion into intended knowledge
3
University of Texas at Dallas
Implicit Knowledge Such knowledge is not directly derivable from
the context of question It belongs either to world knowledge or to
knowledge reported/posted in media articles It belongs to the context of the search It belongs to the domain knowledge
accessible to anyone with expertise in the domain of interest
4
University of Texas at Dallas Two Forms of Question Implicatures
Form 1: Background description Question literal, but there are implicatures between the background description and the question
Example: Q1(Analyst): Recent events in Afhganistan. How have they affected efforts to curb
production on opium in that country?
Implication +translation into simpler question:
How have recent events affected opium production in Afghanistan?
5
University of Texas at Dallas
The answersHow have recent events affected opium production in
Afghanistan? A1: Last fall, as the United States launched its bombing
campaign against the Taliban regime, cash-stripped farmers and warlords eager to make a profit sowed the country’s fields with poppies once again (Source: The Boston Globe; Method: automatic QA)
A2: Since the Taliban regime was ousted and the US-backed regime of Hamid Karzai was installed in Kabul, opium production has risen by one thousand, five hundred tonnes. (Source: Altavista; Method: automatic QA)
6
University of Texas at DallasWhy is it difficult to process Q1?
1) The question does not state which are the recent events in Afghnistan it implies them. Need to generate several intermediary questions for creating the background:
Q1(1): How have recent events affected opium production in
Afghanistan?
Q1(2): How have recent events affected counternarcotics
operations?
2) Q1 has two anaphors: they referring to recent events and that country referring to Aghanistan. Position.
3) Expected answer type: MANNER (action: curb production of opium)
7
University of Texas at Dallas
Source for Semantics WordNet:
synset (control, hold in, hold, contain, check, curb, moderate)
- Paraphrases: production of opium = opium production
What is being done to control opium production in Afghanistan?
A1: The UN drug control programme on Friday welcomed a decision by
Afghanistan’s interim government to offer opium farmers US$250 per destroyed
field.Recent events
8
University of Texas at Dallas Processing background questions
Generate background information Series of questions
Generate context for processing simpler questions. Recognize expected answer from the context of
the background
9
University of Texas at Dallas Case 2: Interpretation relying on Pragmatic Knowledge
Example of Y/N question:
Q2: Will George W. Bush survive the Democrats attacks?1. But they are scaring the Democrats , who are demonstrating palpable fear that in a swing state such as Oregon , where the race between Al Gore and George W. Bush is too close to call , the outsize support for Nader is going to hand the White House to the Republicans...[view]2. Of all the talented people in the Clinton administration , Bush saw fit to keep only two on the job : Dick Clarke , who ran counterterrorism for the National Security Council , and George Tenet , director of the CIA...[view]3. Though there have been plenty of policy disagreements , the Democrats have stood behind their commander in chief , whatever their doubts about his fitness for office or how he attained it...[view]4. " George H.W. Bush attacked Michael Dukakis 's patriotism throughout the 1988 presidential - election campaign...[view]
5. In his speech before hundreds of students , Bradley implored the audience not to vote for Nader...[view]
10
University of Texas at Dallas What is the implication
Survive
George W. Bush Democrat’s attacks
Semantic dependency derivable from the parse
Selectionalconstraints
Dangeroussituation
11
University of Texas at Dallas Two Solutions for Pragmatic Knowledge
1) Rapidly Formatted Knowledge Bases “rapidly” is still time consuming
2) Ad-hoc knowledge on-demand Generate questions when knowledge is needed
Knowledge validation - redundancy on the web - redundancy in categorized textSupport Vector Machines for learning features of ad-hoc
categories Generate Bayesian Networks (probabilistic reasoning) Use Auto-epistemic logic
12
University of Texas at Dallas Pragmatic Knowledge
Auto-epistemic Logic
Organize knowledge + belief operatorsWorld1 World2 Worldn Op1
Op2
Opn
Ad-Hoc Text Categorization
based onSupport Vector
Machines
Generates predominant conceptsand their featured weights
Answer Mining
Question Generation and Processing for Populating possibleworlds
Bayesian Networks
Capture beliefs thatThe epistemic worldsare plausible
13
University of Texas at Dallas Auto-epistemic logic
Attack(Democrat) Opposed-attack(Republican)
Attacks(x,y) actions(x,y) / positions(x,y) / statements(x,y)
Three worlds in auto-epistemic logic (HAEL-style)
Based on generalizations of the mostRepresentative features in the ad-hocCategory involving George W. Bushand the Democrats
Empirical method of generatingAd-hoc categories
14
University of Texas at Dallas Pragmatic knowledge coercion
- Where do we start?
Opposition
Actionsw1
Positionsw2
Statementsw3
Republicans
Democrats
Organized as AutoEpistemic Logic worlds in HAEL–style (Konolidge).+ two operators:Hypothesis – accounts for the best implicaturesStength – accounts for the coercion
Modeled as BaysianNetworks
15
University of Texas at Dallas Populating possible worlds
How do we populate each world?
Actions: select the most general concepts and use it for answer mining
U.S. President has an agenda organized, prioritized actions
Pose to a Q/A system the question:
What items are on the agenda of President Bush?
The most redundant answer on the www (Oct 10, 2002):
War and recession top President Bush’s agenda.
war(a)recession(b) find the context of these items, to enrich the world of actions.
16
University of Texas at Dallas
Text Mining
-For finding the arguments of the newly discovered concepts.
war – has country as semantic category of the dominant argument.
Which country?? Q/A system
1/ Create ad-hoc category WARa) generate topical features from WordNet for the new categoryb) categorize texts with the new features introduced in the SVM model
2/ Detect dominant argument of WAR a) collect windows of 10 words surrounding the most popular feature word b) find the most general semantic category of the most frequent class of words
IRAQ
17
University of Texas at Dallas Validation of Text Mining
Study the usage of FrameNet in deriving pragmatic knowledge by combining extracted information with relational semantics from WN
To date: we have obtained a highly accurate method of classifying any sentence by FrameNet frame (Recent result f-measure = 89%)
18
University of Texas at Dallas Using Pragmatic Knowledge for Deriving Implicatures
Q: Are Democrats for or against war on IRAQ?
The Democrats’ arguments fall closer to the State department’s, which are few and simple:
1) Bush has to “make the case” for war on Iraq. That means prove that Saddam Hussein has chemical or nuclear weapons.
2) Bush must get support for attack from other countries in the world, especially from Europe and the states surrounding Iraq (which has been a failure thus far).
3) Specify the extent of the commitment or resources, troops, money etc. this project is estimated to cost.
4) Involve the people in the decision-making, or better yet Congress.
5) and some are calling for a program of nation-building. The first point is the main one Democrats are repeating, and the rest get less airplay.
19
University of Texas at Dallas Answer mining for implicatures
Q21: Did Bush convince law makers/Congress that US must attack IRAQ?
Q22: What is the US Congress resolution on war on IRAQ?
Q23: How did George W Bush make the case on war on IRAQ?
A4. the past few days. [spacer.gif] [icon0 - print.gif] PRINT [icon0 – discuss.gif] DISCUSSION [icon0 - home.gif] CHINESE [icon0 - sendmail.gif] SEND TO FRIEND [spacer.gif] A choice between war and peace has never been an easy one , but the US Congress swiftly passed a resolution granting President George W. Bush broad authority to act against Iraq after a brief debate over the past few days. This is in sharp contrast with the way the Congress behaved more than a decade ago when the majority of...[view]
20
University of Texas at Dallas
Plausible answers
A4. the past few days. [spacer.gif] [icon0 - print.gif] PRINT [icon0 – discuss.gif] DISCUSSION [icon0 - home.gif] CHINESE [icon0 - sendmail.gif] SEND TO FRIEND [spacer.gif] A choice between war and peace has never been an easy one , but the US Congress swiftly passed a resolution granting President George W. Bush broad authority to act against Iraq after a brief debate over the past few days. This is in sharp contrast with the way the Congress behaved more than a decade ago when the majority of...[view]
Q2: Will George W. Bush survive the Democrats attacks?
Hypothesis(Q22) – was the highest among all possible paths from Q2
Stength(Q22) – measures the plausible quality of the coerced knowledge that
enabled Q22
21
University of Texas at Dallas
Future Plans Model auto-epistemic logic assertions and
operators Quantify the plausibility with Bayesian
Networks Formalize generation of implicatures Study various ways of rejecting implicatures