Abstract: In view of sentence similarity calculation method for intelligent customer service in financial field is studied. Firstly, it reduces the participle errors of Chinese ambiguous words and financial related words by the participle correction model based on part-of-speech. Then, it extracts the semantic features of word level and sentence level and obtains the sentence vectors by the method of word vector and circulatory neural network. In addition, it calculates the discriminative features between sentence vectors by the merge layer. Finally, it obtains the result of sentence similarity calculation by dimension reduction and normalization of the discriminative features. Experiments show that this method has high accuracy and F1 value.

Key words: intelligent customer service, sentence similarity, participle correction, word vector, cyclic neural network

摘要： 针对金融领域中智能客服的句子相似度计算方法进行了研究。利用基于词性的分词纠正模型减少中文歧义词、金融相关词汇的分词错误；通过词向量方法和循环神经网络分别提取词语级和句子级的语义特征，并且得到句子向量；用融合层计算出句子向量间的差异特征；对差异特征进行降维和归一化得到句子相似度计算结果。实验结果表明，该方法具有较高的准确率和[F1]值。

关键词: 智能客服, 句子相似度, 分词纠正, 词向量, 循环神经网络