基于线索词识别和训练集扩展中文问题分类.pdfVIP

① 基于线索词识别和训练集扩展的中文问题分类 张志昌② , 张 宇, 刘 挺, 李 生 (哈尔滨工业大学计算机学院 信息检索研究室, 哈尔滨 150001) 摘 要：针对问题分类的数据稀疏问题，提出了一种以疑问词和焦点词为关键线索的中文事实型问题分类 方法。该方法首先自动识别用户问题中的疑问词和焦点词，若疑问词和焦点词存在，则用最近邻模型进行 分类，而对没有用最近邻方法分类的其他问题，则用支持向量机（SVM）模型进行分类。训练 SVM 模型时， 从 Web 上自动获取新问题来对训练集进行扩展，最近邻方法只利用线索词词义距离进行类别判断。实验表 明，这种按照问题结构的不同而选择不同分类器的方法，在性能上要优于单一分类方法；词义距离的应用 和训练集自动扩展改善了训练数据的稀疏，提高了分类性能。 关键词：问题分类；焦点词；词义距离；训练集扩展 中图法分类号： TP391 文献标识码： A 文章编号：xxxx-xxxx(xxxx)xx-xxxx-xx Chinese Question Classification Based on Identification of Cue Words and Extension of Training Set ZHANG Zhichang, ZHANG Yu, LIU Ting, LI Sheng (Information Retrieval Laboratory of Computer Science Technology School, Harbin Institute of Technology, Harbin 150001, China) Abstract: Focusing on data sparseness problem in question classification, an approach for Chinese factoid questions was proposed, which used interrogative and focus words as key cues. The Nearest Neighbor model was used to classify the questions in which cue words could be identified automatically, then SVM was used for other questions. The training set of SVM was extended automatically with questions mined from Web while the Nearest Neighbor model only used the sense distance of the cue words as feature. The experimental results show that the approach, selecting different classifier according to question structure, outperforms single classification model, and the training data sparseness is alleviated using word sense distance and extension of training set, thus classification performance is improved. Keywords: question classification; focus word; word sense distance; extension of training set 0 引言 问答技术在信息检索领域中已成为重要的研究方向，其目标是根据用户用自然语言提出的问题，从大 规模文档集合中提取出对该问题的正确答案。现有的问