Abstract:

Comparative text is very important for competitive products analysis, but there are few researches on the classification of comparative text in the Q&A field. Aiming at the characteristics of rich information and concentrated topics in Q&A texts, this paper proposes a comparative text classification method based on topic feature and keyword feature expansion. Based on the pretrained topic model, the topic probability distribution of the Q&A text is inferred as its topic feature. In view of the keyword information loss caused by vector concatenation and summation, GRU-autoencoder is designed to realize feature extraction, and the encoder output is used as the keyword feature of Q&A text. Integrating the topic information and keyword semantics, the comparative text features are constructed from the perspectives of linguistics, product, sentiment, social, topic and keyword, then the Q&A text is classified by using various classifiers. The experimental results show that the constructed features are effective and the effect of the classification are better.

Key words: topic model, autoencoder, feature expansion, comparative text classification

摘要：

比较文本对于企业竞争产品分析至关重要，但目前面向问答领域的比较文本分类研究较少。针对问答文本中比较信息丰富、主题集中的特点，提出了基于主题特征和关键词特征扩展的比较文本分类方法。通过预训练主题模型，推断问答文本的主题概率分布作为其主题特征；针对向量拼接、求和导致关键词信息流失的问题，设计GRU自编码器实现关键词向量特征提取。综合文本主题信息和关键词语义，从语言、产品、情感、社交、主题、关键词角度构建比较文本分类特征，最后使用多种分类器对问答文本进行分类。实验结果表明，构建的特征行之有效，比较文本分类效果较好。

关键词: 主题模型, 自编码器, 特征扩展, 比较文本分类