Some Challenges of Advanced Question-Answering: an Experiment with How-to Questions

Some Challenges of Advanced Question-Answering: an Experiment with How-to Questions * Patrick Saint-Dizier IRIT-CNRS, 118, route de Narbonne 31062 Toulouse, France stdizier@irit.fr Abstract. This paper is a contribution to text semantics processing and its application to advanced question-answering where a significant portion of a well-formed text is required as a response. We focus on procedural texts of various domains, and show how titles, instructions, instructional compounds and arguments can be extracted. Keywords: text semantics, question answering. 1. Introduction Question answering (QA) is an area that operates on top of search engines (such as Google, Exalead, Yahoo, etc.) in order to provide users with more accurate and elaborated responses where search engine response outputs remain punctual, difficult to understand, and sometimes incoherent. QA builds on top of search engine responses via language processing tools, reasoning mechanisms and response production techniques. QA has been recognized as an essential component of man-machine communication since it greatly improves communication with the Web or textual databases, allowing users to communicate in their own language, in order to get much more precise, accurate and user-tailored responses. Factoid questions (questions about facts, e.g.: what is the capital of X ?) have been largely investigated (witness, e.g. the TREC competition), but the other types of questions, frequently encountered such as : procedural (how to), causal (why), evaluative and comparative questions (what is the cheapest air ticket to go to Cebu from Singapore ?) have received little attention so far. However, they correspond to the majority of the 'real' questions asked either by the large public or by professionals to the web or to textual databases. Questions cover a large number of everyday life (health, employment, administrative life, tourism, etc.) as well as technical areas (investments, competitive intelligence, etc.). Complex questions require language processing (to process the question and on the outputs of search engines), reasoning (e.g. data fusion, incoherence solving, summarizing, etc.) and response production in natural language (in a way accessible and user-friendly). These complex questions are in fact those users ask most of the time, these users being professionals or from the large public. Complex questions are traditionally classified according to the type of response which is induced, e.g.: - definition questions, * Acknowledgments: we are grateful to a number of MA students who contributed to this work, among which: Estelle DELPECH, Lionel FONTAN, Isabelle DAUTRICHE and Clementine ADAM. We also thank the ANR-RNTL French program for supporting part of this research. Copyright 2008 by Patrick Saint-Dizier. 65 22nd Pacific Asia Conference on Language, Information and Computation, pages 65–73 - how-to questions for procedures, - why questions for causes/consequence (possibly involving chains of events), - evaluative and comparative questions (asking for comparisons between two or more elements such as fares). Considering the technology which is required, it is clear that search engines will not be able in the near future or even in the mid-term future to correctly process these questions due to the complexity of the task, the need of some domain knowledge and the …

Content truncated to 500 words