Exploring Syntactic Relation Patterns for Question Answering

R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 507 – 518, 2005. © Springer-Verlag Berlin Heidelberg 2005 Exploring Syntactic Relation Patterns for Question Answering Dan Shen1,2, Geert-Jan M. Kruijff 1, and Dietrich Klakow2 1 Department of Computational Linguistics, Saarland University, Building 17, Postfach 15 11 50, 66041 Saarbruecken, Germany {dshen, gj}@coli.uni-sb.de 2 Lehrstuhl Sprach Signal Verarbeitung,Saarland University, Building 17, Postfach 15 11 50, 66041 Saarbruecken, Germany {dietrich.klakow}@lsv.uni-saarland.de Abstract. In this paper, we explore the syntactic relation patterns for open- domain factoid question answering. We propose a pattern extraction method to extract the various relations between the proper answers and different types of question words, including target words, head words, subject words and verbs, from syntactic trees. We further propose a QA-specific tree kernel to partially match the syntactic relation patterns. It makes the more tolerant matching be- tween two patterns and helps to solve the data sparseness problem. Lastly, we incorporate the patterns into a Maximum Entropy Model to rank the answer candidates. The experiment on TREC questions shows that the syntactic rela- tion patterns help to improve the performance by 6.91 MRR based on the com- mon features. 1 Introduction Question answering is to find answers for open-domain natural language questions in a large document collection. A typical QA system usually consists of three basic modules: 1. Question Processing (QP) Module, which finds some useful information from questions, such as expected answer type and key words; 2. Information Retrieval (IR) Module, which searches a document collection to retrieve a set of relevant sen- tences using the key words; 3. Answer Extraction (AE) Module, which analyzes the relevant sentences using the information provided by the QP module and identify the proper answer. In this paper, we will focus on the AE module. In order to find the answers, some evidences, such as expected answer types and surface text patterns, are extracted from answer sentences and incorporated in the AE module using a pipelined structure, a scoring function or some statistical-based meth- ods. However, the evidences extracted from plain texts are not sufficient to identify a proper answer. For examples, for “Q1910: What are pennies made of?”, the expected answer type is unknown; for “Q21: Who was the first American in space?”, the sur- face patterns may not detect the long-distance relations between the question key phrase “the first American in space” and the answer “Alan Shepard” in “… that car- ried Alan Shepard on a 15 - minute suborbital flight in 1961 , making him the first 508 D. Shen, G.-J.M. Kruijff, and D. Klakow American in space.” To solve these problems, more evidences need to be extracted from the more complex data representations, such as parse trees. In this paper, we explore the syntactic relation patterns (SRP) for the AE module. An SRP is defined as a kind of relation between a question word and an answer can- didate in the syntactic tree. Different from the textual patterns, the SRPs capture the relations based on the sentence …

Content truncated to 500 words