Context

Question-answering is a very challenging area with still many open challenges. Our group is oriented towards advanced forms of question answering that require explicit linguistic processing and the integration of knowledge and dedicated forms of reasoning to produce answers. Moreover, answers are seldom simple and often require the use of natural language generation technology to produce well-organized, well-planned and cooperative responses.

Overview

We address in particular the following types of questions

  • How-to Questions: where the response a well-formed text, in general a procedure. This type of question requires adapted forms of indexing (on titles representing goals and sub-goals), the analysis of procedural texts so that the appropriate portion of the procedure can be given to the user, with appropriate pre-requisites, warnings, advices, comments, etc. Additional research related to the TextCoop project described above.
  • Evaluative and comparative questions, where the challenge is to identify and develop the semantics of comparative expressions in context (which cities are the most kid-friendly in Asia?, what are the most innovative works in this lab?) and to elaborate means (using set theoretical tools, reasoning) to compute the response from data, where the response is not explicitly given, but must be constructed. We develop a particular application in business intelligence in cooperation with ESC Toulouse and De La Salle University, Manila, the Philippines.
  • Why-questions involving chains of causes and consequences. This work is realized in conjunction with Thai partners (Kasetsart university) and applied to agriculture. Challenges of this type of question are also very important, e.g. identifying appropriate and related chains of events in a text (possibly not contiguous) leading to a certain conclusion, representing the semantics of the question which may be quite different from the response documents, developing question-document matching techniques.
  • Opinion questions applied to the analysis of journal editorials from different origins on precise events. The aim here is to gather contradicting opinions on a fact, with arguments, and developments realized via rhetorical relations (elaboration, reformulation, reinforcement, example, etc.). The ultimate goal is to produce, over time, a map of opinions on a certain fact with its evolution. This work is realized in cooperation with Kathmandu University, Nepal.
  • Introducing cooperativity in responses: in a large number of situations the question has no direct response: it is therefore necessary for the system to have a cooperative behavior to get ‘approximate' responses and to produce explanations to the user.

Contributors

Main publications

  • Mukda Suktarachan, Patrick Saint-Dizier, Asanee Kawtrakul, Using Lexical Semantics for Question-Answering in e-Farming, WCCA conference, Reno, May 2009.
  • Bal Krishna Bal, Patrick Saint-Dizier, Towards and Analysis of Argumentation Structure and the Strength of Arguments in News Editorials, AISB Symposium on Persuasive Technologies, Edinburgh, April 2009.
  • Patrick Saint-Dizier, Some Challenges of Advanced Question-Answering, invited talk, PACLIC Conference, Springer Verlag lecture notes, Cebu, the Philippines, November 2008.
  • Chaveevan Pechsiri, Asanee Kawtrakul, Elixabete Murguia, Patrick Saint-Dizier. Mining Causality from Texts to Construct Explanation Knowledge Expertise. in : ECTI Transactions, ECTI Association, Bangkok, V. 1 N. 4, february 2007.