- Количество слайдов: 42
Answering complex questions and performing deep reasoning in advance QA systems: ARDA AQUAINT Program Phase 2 Chitta Baral Arizona State university
Participants and other students l Arizona State University – – l Texas Tech University – – l PI: Chitta Baral Chitta’s student participants: Luis Tari, Jicheng Zhao, Hiro Takahashi, Saadat Anwar, Ryan Weddle (during summer), Nam Tran, Xin Zhang, Piyun Chang (during summer) Chitta’s other students: Le-Chi Tuan Other students: Deepthi Chidambaram, Toufeeq Ahmed PI: Michael Gelfond Student Participants: Marcello Balduccini, Greg Gelfond Monmouth University – PI: Richard Scherl
AQUAINT Program goals (from BAA 03 -06 -FH) l l … seeks proposals for innovative, creative and high-risk research, which will continue to advance the state-of-the-art in technologies and methods for advanced, automated question answering. Phase 1 Research goals and accomplishments: focused on the following functional components and enabling technologies – – 1. Question understanding and interpretation 2. Determining the answer 3. Formulating and presenting the answer 4. Cross-cutting/Enabling/Enhancing technologies that directly and materially support the goals of the AQUAINT program and one or more of the areas 1 -3 listed above.
AQUAINT Program goals (from BAA 03 -06 -FH) -- cont l Phase 2 Research Goals: In addition to pursuing the goals identified in Phase 1 of the program, Phase 2 will encourage efforts, ehich focus on the following challenges – 1. Question answering as part of a larger informationgathering process (2 implications) l l – – – Increasing complexity of questions Synthesis of Information found in multiple data sources 2. Accessing, Retrieving and Integrating diverse data sources 3. Exploring boundaries/combinations of knowledge-based, statistical and linguisitic approaches to question answering 4. Evaluating, Validating and Presenting an answer
Some focal points of our project: excerpts from BAA 03 -06 -FH l Increasing Complexity of Questions: In addition to the more factually based, who, what, when, where type of questions that today’s state of the art Q & A systems tackle, the ultimate, advanced Q & A systems must be able to successfully respond to the far more complex why and how types of questions. These complex questions will likely involve judgment terms involving intent, motive, meaning, reason, purpose, aim, objective, implications, etc. or the questions might require the advanced Q & A system to compare, contrast, examine, inspect, match, size up, weigh, etc. two or more different yet related entities, objects or positions. And finally the questions asked of this ultimate system will at times tend to be somewhat vague, open-ended and abstract.
Some focal points of our project as excerpted from BAA 03 -06 -FH (cont. ) l l l The advanced Q & A system needs to recognize when it can not find or does not know the answer to the original question. Clearly, systems … perform deep reasoning and … complex chains of inference. Although the focus of ARDA’s AQUAINT program is to tackle and research unsolved technical problems, it is important to remember that the ultimate goal of the program is to develop Q & A systems which can be made available as automated tools to the intelligence analyst.
Focus of our proposal (from our statement of work) l Develop various component elements focused on the following issues – – – Answering increasingly complex questions Figuring out whether a particular question can be answered with the given information and if not, either giving qualified or tentative answers or … Developing ways to meaningfully present answers
Further elaboration of our goals: take QA to another level, beyond simple querying l l l Answer hypothetical queries, narrative queries, counterfactual queries, planning queries etc. Reasoning with incomplete information, defaults, normative statements, etc. Formulating deep reasoning notions such as when is a behavior or event abnormal or suspicious, when a statement is a lie, what is an explanation, what is a diagnosis, what is a cause, etc.
QUERIES Prediction, explanation, planning, cause, counterfactual, etc.
Queries and Answers l Answering queries with respect to databases: various query languages – – l Natural language queries – l Relational databases: SQL 3 Object-Oriented Databases: OQL Web databases, XML Databases: XML-QL Prolog queries Often translated to one of the above Complex Queries! – – Need knowledge beyond that is present in the given data (or text) to answer. Need reasoning mechanisms that can not be expressed in standard database query languages or classical logics.
Complex Query example – predictive query l l Text/Data: John is at home in Boston and has not bought a ticket to Paris yet. Query: – – l What happens if John tries to fly to Paris? What happens if John buys a ticket to Paris and then tries to fly to Paris? Missing knowledge: – – When can one fly? What is the result of flying?
Complex Query example: explanation query l l Text/Data: On Dec 10 th John is at home in Boston and does not have a ticket to Paris yet. On Dec 11 th he is in Paris. Query: – Explain what might have happened in between. l Bought a ticket; gone to the Boston airport; taken a flight to Paris.
Complex Query Example: planning query l l Text/Data: On Dec 10 th John is at home in Boston and does not have a ticket to Paris yet. Query: What does John need to do to be in Paris on Dec 11 th. – He needs to buy the ticket || get to the airport; fly to Paris.
Complex Query Example: Counterfactual Query l l Text/Data: On Dec 10 th John is at home in Boston. He made a plan to get to Paris by Dec 11 th. He then bought a ticket. But on his way to the airport he got stuck in the traffic. He did not make it to the flight. Query: What if John had not gotten stuck in the traffic?
Complex Query Example: query about narratives l l Text/Data: John, who always carries his laptop with him, took a flight from Boston to Paris on the morning of Dec 11 th. Queries: – – Where is John on the evening of Dec 11 th? In which city is John’s laptop that evening?
Complex Query Example: Causal queries l l Text/Data: On Dec 10 th John is at home in Boston. He made a plan to get to Paris by Dec 11 th. He then bought a ticket. But on his way to the airport he got stuck in the traffic. He reached the airport late and his flight had left. Queries: – What are the causes of John missing the flight?
Complex Query Example: Unusual behavior l l John flew from Boston to Paris. He did not check in any luggage in Boston. When he got out of the plane in CDG he did not have anything in his hand. Was there anything unusual about John’s behavior when he checked in? – – Need information on normal behavior of people who check in for an international flight Normal inertia with respect to hand luggage (from checking in to getting out of the plane)
Our approach and progress
Basic thesis l l l The documents on which Q & A is to be based often does not contain the general knowledge necessary to answer deep questions. This knowledge has to be written for a system to be able to do deep reasoning. Basic questions: – – How to write this knowledge (in which language) How to do various kinds of deep reasoning with this knowledge together with information embedded in the given documents?
Starting Point l l Past research in knowledge representation and reasoning. The book on the left. Initial article was by Gelfond and Lifschitz. (Rannked 17 th in the most cited list http: //citeseer. ist. psu. edu/sourc e. html) http: //citeseer. ist. psu. edu/allcite d. html – – – Gelfond (268) Baral (2757) Scherl (4948)
Post-contract plan of action l l 1. Use the existing knowledge representation theory and systems to do deep reasoning 2. Enhance theory 3. Enhance systems 4. Prepare for bridging with other projects to lead to an end-to-end system
Work in progress and today’s agenda l Morning session – – l (1) Progress on using existing theory and systems to encode common-sense knowledge and use it to answer difficult queries. (Richard Scherl) (3) Adding a GUI to the Smodels reasoning system (Hiro Takahashi) (2, 3) CR-Prolog (Marcello Balduccini) (2) Enhancing Ans. Prolog to reason with probabilities (Chitta Baral) Afternoon session – – – (4) A simple QA system (Piyun Chang) (4) A Text extraction system used with respect to Bio-medical texts (Deepthi Chidambaram, Toufeeq Ahmed -- student’s of my colleague Hasan Davulcu) (4) NLP module to translate English questions to our representation (Richard Scherl) (2) Goal Language (Jicheng Zhao) *if there is time (2) Modules (Luis Tari) *if there is time Overview of other work in Chitta’s Lab.
NEXT Richard Scherl
Our approach to answer such queries l Develop various knowledge modules in an appropriate knowledge representation and reasoning language. – – – – – Travel module (includes flying, driving) Geography Module Time module Reasoning about actions module Planning module Explanation module Counterfactual module Cause finding module Most of the above modules with defaults and exceptions.
Knowledge Representation and Reasoning: Ans. Prolog
What properties should an appropriate KR & R language have l l l l Should be non-monotonic. So that the system can revise its earlier conclusion in light of new information. Should have the ability to represent normative statements, exceptions, and default statements, and should be able to reason with them. Should be expressive enough to express and answer problem solving queries such as planning queries, counterfactual queries, explanation queries and diagnostic queries. Should have a simple and intuitive syntax so that domain experts (who may be non-computer scientists) can express knowledge using it. Should have enough existing research (or building block results) about this language so that one does not have to start from scratch. Should have interpreters or implementation of the language so that one can indeed represent and reason in this language. (I. e. , it should not be just a theoretical language. ) Should have existing applications that have been built on this language so as to demonstrate the feasibility that applications can be indeed built using this language.
Ans. Prolog – a suitable knowledge representation language l Ans. Prolog – Programming in logic with answer sets – l Language (and semantics) was first introduced in the paper ``The Stable Model Semantics For Logic Programming - Gelfond, Lifschitz (1988)’’, among the most cited source documents in the Cite. Seer database. http: //citeseer. ist. psu. edu/source. html Syntax: Set of statements of the form: A 0 or … or Ak B 1, …, Bm, not C 1, … not Cn. l Intuitive meaning of the above statement: – l If B 1, …, Bm is known to be true and C 1, …, Cn can be assumed to be false then at least one of A 0 , …, Ak must be true. It satisfies all the properties mentioned in the previous slide (and much more)! – Details in my Book ``Knowledge Representation, Reasoning and Declarative Problem Solving’’. Cambridge University Press, 2003.
Ans. Prolog vs Prolog l Differences: – – l Prolog is sensitive to ordering of rules and ordering of literals in the body of rules. Inappropriate ordering leads to infinite loops. (Thus Prolog is not declarative; hence not a knowledge representation language) Prolog stumbles with recursion through negation No disjunction in the head (less power) Similarities: For certain subclasses of Ans. Prolog can be thought of as a top-down engine.
Ans. Prolog vs other KR & R languages l l Ans. Prolog has a simple syntax and semantics Syntax has structure that allows defining sub-classes More expressive than propositional and first-order logic; propositional Ans. Prolog is as expressive as default logic. Yet much simpler. It has a very large body of support structure (theoretical results as well as implementations) among the various knowledge representation languages – l … Description logic comes close. But its focus is somewhat narrow, mostly to represent and reason about ontologies.
Illustration of Complex Query Answering John flying to Baghdad to meet Bob example.
The extracted text and the queries l Extracted Text – l John spent Dec 10 in Paris and took a plane to Baghdad the next morning. He was planning to meet Bob who was waiting for him there. Queries – – Q 1: Was John in the Middle East in mid. December? Q 2: If so, did he meet Bob in the Middle East in mid-December?
Required background and commonsense knowledge l Knowledge about geographical objects and their hierarchy. (M 1) – – l Knowledge about travel events. (M 2) – – l If someone is in a city then she is in the country where the city is in and so on. Executability conditions and effect of travel events Inertia Duration of flying Knowledge about time units. (M 3) – l Baghdad is a city in Iraq is a country in the middle east region. … A city in a country in a region is a city in that region. Relation between various time granularities Knowledge about planned events, meeting events. (M 4) – – People normally follow through their plans Executability condition of meeting events
M 1: The geography Module l List of places – – – l Relation between places – – – l is(baghdad, city). is(iraq, country). . in(baghdad, iraq). in(iraq, middle_east). in(paris, france). in(france, western_europe). in(western_europe, europe). . Transitive closure in(P 1, P 3) in(P 1, P 2), in(P 2, P 3). l Completeness assumption about `in’ -in(P 1, P 2) not in(P 1, P 2)
M 2: The traveling module l Based on theory of dynamic systems – Views world as a transition diagram l l l Various types of traveling events – – l instance_of(fly, travel). instance_of(drive, travel). … Generic description of John flying to Baghdad – – l States are labeled by fluents Arcs labeled by actions event(a 1). type(a 1, fly). actor(a 1, john). destination(a 1, baghdad). Actual event is recorded as – occurs(a 1, i)
M 2: The traveling module (cont. ) l Representation of transition Diagram – State Constraints loc(P 2, X, T) loc(P 1, X, T), in(P 2, P 1). disjoint(P 1, P 2) -in(P 1, P 2), -in(P 2, P 1), neq(P 1, P 2). -loc(P 2, X, T) loc(P 1, X, T), disjoint(P 1, P 2). – Causal Laws loc(P, X, T+1) occurs(E, T), type(E, travel), actor(E, X), destination(E, P), -interference(E, T) not interference(E, T). – Executability Conditions -occurs(E, T) cond(T). – Inertia Rules (frame axioms) loc(P, X, T+1) loc(P, X, T), not -loc(P, X, T+1) -loc(P, X, T), not loc(P, X, T+1).
Reasoning with M 1 and M 2 l Given – – – l loc(paris, john, 0). loc(baghdad, bob, 0). occurs(a 1, 0). And with M 1 and M 2 Ans. Prolog can conclude – – loc(baghdad, john, 1), loc(baghdad, bob, 1), loc(middle_east, john, 1), -loc(paris, john, 1)
M 3: Time and durations l Duration of actions (additional ones needed for month etc. ) time(T+1, day, D) occurs(E, T), type(E, fly), time(T, day, D), not -time(T+1, day, D). l Basic measuring units – l day(1. . 31). month(1. . 12). part(start). part(end). part(middle). Rules translating between one granularity to another time(T, part, middle) time(T, d, D), 10 < D < 20. time(T, season, summer) time(T, month, M), 5 < M < 9. l Missing elements from the module – – next(date(10, 12, 03), date(11, 12, 03)). next(date(31, 12, 03), date(1, 1, 04)).
Reasoning with M 1, M 2 and M 3 l Given information about John’s flight – – – l The query Q 1 – l loc(paris, john, 0). loc(baghdad, bob, 0). occurs(a 1, 0). time(0, day, 11). time(0, month, 12). ? loc(middle_east, john, T), time(T, month, 12), time(T, part, middle). Ans. Prolog gives the correct answer: yes with T = 1.
M 4: planning to meet and meeting l Describing the event meet – – – l event(a 2). type(a 2, meet). actor(a 2, john). actor(a 2, bob). place(a 2, baghdad). Executability conditions of the meeting event -occurs(E, T) type(E, meet), actor(E, X), place(E, P), -loc(P, X, T). l l Planned meeting: planned(a 2, 1). Planned actions and their occurrence: ``People normally follow their plans’’ occurs(E, T) planned(E, T), not -occurs(E). l People persist with their plans until it happens planned(E, T+1) planned(E, T), -occurs(E, T). l Second query – l ? occurs(E, T), type(E, meet), actor(E, john), actor(E, bob), loc(middle_east, john, T), time(T, month, 12), time(T, part, middle). Answer: Yes.
Conclusion l l Answering complex queries need a lot of knowledge and reasoning rules that are not present in the text or data. These reasoning rules and knowledge need to be encoded as modules in an appropriate knowledge representation and reasoning language.
Ongoing and Future Work l Further development of Modules (examples) – – – l Travel duration Time period representation issues (eg. time zones) Dealing with the case when a planned event fails Further development of the Ans. Prolog language – Not good when dealing with time or similar features that result in large instantiations. l – Taking advantage of Prolog execution engine when necessary Necessity of set notations, aggregates etc.
Acknowledgements l l Steve Maiorano, Jean-Michel Pomarede Ryan Weddle, Jicheng Zhao, Saadat Anwar, Luis Tari (all from ASU), Greg Gelfond (from Texas Tech University)