深度文本匹配综述

摘要

自然语言理解的许多任务,例如信息检索、自动问答、机器翻译、对话系统、复述问题等等,都可以抽象成文本匹配问题.过去研究文本匹配主要集中在人工定义特征之上的关系学习,模型的效果很依赖特征的设计.最近深度学习自动从原始数据学习特征的思想也影响着文本匹配领域,大量基于深度学习的文本匹配方法被提出,作者称这类模型为深度文本匹配模型.相比于传统方法,深度文本匹配模型能够从大量的样本中自动提取出词语之间的关系,并能结合短语匹配中的结构信息和文本匹配的层次化特性,更精细地描述文本匹配问题.根据特征提取的不同结构,深度文本匹配模型可以分为3类:基于单语义文档表达的深度学习模型、基于多语义文档表达的深度学习模型和直接建模匹配模式的深度学习模型.从文本交互的角度,这3类模型具有递进的关系,并且对于不同的应用,具有各自性能上的优缺点.该文在复述问题、自动问答和信息检索3个任务上的经典数据集上对深度文本匹配模型进行了实验,比较并详细分析了各类模型的优缺点.最后该文对深度文本模型未来发展的若干问题进行了讨论和分析.%Many problems in natural language processing,such as information retrieval,question answering,machine translation,dialog system,paraphrase identification and so on,can be treated as a problem of text matching.The past researches on text matching focused on defining artificial features and learning relation between two text features,thus the performance of the text matching model heavily relies on the features designing.Recently,affected by the idea of automatically feature extraction in deep learning,many text matching models based on deep learning,namely Deep Text Matching model,have been proposed.Comparing to the traditional methods,Deep Text Matching models can automatically learn relations among words from big data and make use of the information from phrase patterns and text hierarchical structures.Considering the different structures of Deep Text Matching models,we divide them into three categories:Single semantic document representation based deep matching model,Multiple semantic document representation based deep matching model and Matching pattern based deep matching model.We can see the progressive relationship among three kinds of models in modelling the interaction of texts,while which have their own merits and defects based on a specific task.Experiments were carried out on the typical datasets of paraphrase identification,question answering and information retrieval.We compare and explain the different performance of three kinds of deep text matching models.Finally,we give the key challenges and the future outlooks of the deep text matching models.