Abstract: In view of players’ professional and colloquial way of querying in game customer service scenarios, this paper presents a sentence similarity model that takes into account synonymous substitutions, weights, sentence length, word order and other factors with semantic word vectors being established using the deep learning tool word2vec. Based on this model, the drawbacks of both dominance of majority classes and high computational cost associated with KNN classification algorithm are improved by pre-classification and re-defining classification rules. Furthermore, this paper implements an automatic question-answer system based on text classification for the game customer service scenarios. Experimental results show that this system has higher accuracy and efficiency of queries classification.

Key words: word2vec, sentence similarity, text classification, automatic question-answer, natural language processing

摘要： 针对游戏客服场景中玩家领域化、口语化的提问方式，应用深度学习工具word2vec建立带有语义的词的向量表示，设计了一种利用词向量距离，结合同义词替换、权重、句子长度、词序等因素的句子相似度计算模型。在该模型基础上，通过预分类、重定义分类规则，对KNN分类算法的大类占优、全局匹配计算代价高等问题进行改进，实现了一种基于文本分类的面向游戏客服场景的自动问答系统。实验结果表明，该系统具有较高的问题分类准确率和分类效率。

关键词: word2vec, 句子相似度, 文本分类, 自动问答, 自然语言处理