Using a Trie-based Structure for Question Analysis

Using a Trie-based Structure for Question Analysis Luiz Augusto Sangoi Pizzato Centre for Language Technology Macquarie University 2109 Sydney, Australia pizzato@ics.mq.edu.au http://www.clt.mq.edu.au Abstract This paper presents an approach for question analysis that defines the question subject and its required answer type by building a trie- based structure from a set of question patterns. The question analysis consists of comparing the question tokens with the path of nodes in the trie. A look-ahead process solve the mismatches of unknown words by assigning a entity-type or semantically linking them with other question words. The developed approach is evaluated using different datasets showing that its perfor- mance is comparable with state-of-the-art sys- tems. 1 Introduction When a question is presented to a person, or even to an automatic system, the first task, in order to provide an answer, is to understand the question. The question analysis process may not be very clear for people when answering questions, however for an automatic question answering (QA) system it plays a crucial role. Acquiring the information embedded in a question is the primary task that allows the system to execute the right commands in order to provide the correct answer to it. According to Moldovan et al. (2003), when the question analysis fails, it is hard or almost impossible for a QA system to perform its task. The impor- tance of the question analysis is very clear in the system of Moldovan et al. (2003) since this task is performed by 5 of the 10 modules that compose their system. The most common approach for analysing questions is to divide the task into two parts: Finding the question expected answer type, and finding the question focus. Many systems (Moll´a-Aliod, 2003; Chen et al., 2001; Hovy et al., 2000) use a set of hand- crafted rules for finding the expected answer type (EAT). Normally the rules are written as regular expressions (RE), while the task of find- ing the EAT consists of matching questions and REs. Every RE will have an associated EAT that will be assigned to a question if it matches its pattern. For the task of finding the question focus, the simplest approach is to discard every stopword on the question and to consider the remaining terms as the focus representation. In the approach described in this paper, the EAT and the question focus are defined using a trie-based structure built from a manually anno- tated corpus of questions. The structure stores the answer type in every trie node and uses the question words or entity types to link the nodes. The question analysis method was evaluated over an annotated set of question of an acad- emic domain, over the annotated TREC-2003 questions and over the 6,000 questions of the training/testing set of question of Li and Roth (2002) showing promising results. This paper addresses a technique used to analyse natural language (NL) questions and its evaluation. Section 2 describes the technique, while Section 3 presents its evaluation. In Sec- tion 4 some related …

Content truncated to 500 words