Abstract: How to understand a user’s intention according to the context in multi-round chatting robots is a key and difficult issue. Most existing questions-understanding methods are for a single sentence, and focus on the understanding of a sentence structure. It is an urgent problem to be solved that how to identify the current user’s intention according to the context instead of analyzing each single round individually, which can make the dialogue have fine-grained understanding ability in a continuous context. In order to solve the above problems, this paper presents a method of multi-intention classification of natural language questions based on deep learning. The user intents include 13 categories, such as chatting, music, news, arithmetic, catering, booking, weather, service categories and so on. Firstly, this paper uses the technology of natural language processing to process and analyze multi-round dialogues, identifies the key words therein, and then constructs the dichotomous and multi-category deep learning model to learn contexts by using the deep learning method and the hierarchical classification technology, they are both used to identify user’s intention. Experiments show that the accuracy rate of user intention recognition with constructed deep learning model is 94.81% and 93.49% respectively. Therefore, the proposed method can basically solve the problem of understanding a user’s intention with natural question language.

Key words: intent identification, natural language processing, deep learning, classification

摘要： 在聊天机器人多轮对话中如何根据上下文理解用户的意图是多轮交互中的一个重点问题，也是一个难点问题。现有的问句理解方法大多是针对单句的，且侧重于某种句式结构的理解。如何根据上下文语境对当前用户的意图进行识别，而不仅仅是针对单轮进行一个个分析，使得对话在一个连续的语境下具备细粒度的理解能力，是一个亟待解决的问题。针对以上问题，提出了一种基于深度学习的自然语言问句多意图分类方法，其中涉及到的用户意图包含闲聊类、音乐类、新闻类、算术类、餐饮类、订票类、天气类、服务类等13类。首先使用自然语言处理的相关技术对多轮对话进行处理分析，识别出其中的关键词，然后使用深度学习方法和分层分类技术构建了二分类和多分类深度学习模型，学习上下文语境和语义关系，共同对用户意图进行识别。通过实验证明了构建的深度学习模型对用户意图识别的准确率分别为94.81%、93.49%。因此，所提方法基本能够解决自然语言问句意图识别的问题。

关键词: 意图识别, 自然语言处理, 深度学习, 分类