Classifying What-Type Questions by Head Noun Tagging

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 481–488 Manchester, August 2008 Classifying What-type Questions by Head Noun Tagging Fangtao Li, Xian Zhang, Jinhui Yuan, Xiaoyan Zhu State Key Laboratory on Intelligent Technology and Systems Tsinghua National Laboratory for Information Science and Technology Department of Computer Sci. and Tech., Tsinghua University, Beijing 100084, China zxy-dcs@tsinghua.edu.cn Abstract Classifying what-type questions into proper semantic categories is found more challenging than classifying other types in question answering systems. In this paper, we propose to classify what-type questions by head noun tagging. The ap- proach highlights the role of head nouns as the category discriminator of what- type questions. To reduce the semantic ambiguities of head noun, we integrate local syntactic feature, semantic feature and category dependency among adjacent nouns with Conditional Random Fields (CRFs). Experiments on standard ques- tion classification data set show that the approach achieves state-of-the-art per- formances. 1 Introduction Question classification is a crucial component of modern question answering system. It classifies questions into several semantic categories which indicate the expected semantic type of answers to the questions. The semantic category helps to filter out irrelevant answer candidates, and de- termine the answer selection strategies. 1 The widely used question category criteria is a two-layered taxonomy developed by Li and Roth (2002) from UIUC. The hierarchy contains 6 coarse classes and 50 fine classes as shown in Table 1. In this paper, we focus on fine-category classification. Each fine category will be denoted as “Coarse:fine”, such as “HUM:individual”. A what-type question is defined as the one whose question word is “what”, “which”, “name” or “list”. It is a dominant type in ques- tion answering system. Li and Roth (2006) find © 2008. Licensed under the Creative Commons Attri- bution-Noncommercial-Share Alike 3.0 Unported license (http://creativecommons.org/licenses/by-nc- sa/3.0/). Some rights reserved. Coarse Fine ABBR abbreviation, expression DESC definition, description, manner, reason ENTY animal, body, color, creation, currency, dis- ease/medicine, event, food, instrument, language, letter, other, plant, product, religion, sport, sub- stance, symbol, technique, term, vehicle, word HUM description, group, individual, title LOC city, country, mountain, other, state NUM code, count, date, distance, money, order, other, percent, period, speed, temperature, size, weight Table 1. Question Ontology that the distribution of what-type questions over the semantic classes is quite diverse, and they are more difficult to be classified than other types. Table 2 shows the classification accuracies of each question word in UIUC data set using Sup- port Vector Machine (SVM) with unigram fea- tures. What-type questions account for more than 70 percent in the data set, but the classification accuracy of this type only achieves 75.50%. In this experiment, 90.53% (86 over 95) of the er- rors are generated by what-type questions. Due to its challenge, this paper focuses on what-type question classification. Total Wrong Accuracy What-type 351 86 75.50% Where 26 2 92.31% When 26 0 100.0% Who 47 3 93.62% How 46 4 91.30% Why 4 0 100.0% Total 500 95 81.00% Table 2. Classification performance for each …

Content truncated to 500 words