Convolutional Neural Network Architectures for Matching NaturalLanguage Sentences

匹配自然语言句子的卷积神经网络结构

Abstract

Semantic matchingis of central importance to many natural language tasks [2, 28]. A successfulmatching algorithm needs to adequately model the internal structures oflanguage objects and the interaction between them. As a step toward this goal,we propose convolutional neural network models for matching two sentences, byadapting the convolutional strategy in vision and speech. The proposed modelsnot only nicely represent the hierarchical structures of sentences with theirlayerby-layer composition and pooling, but also capture the rich matchingpatterns at different levels. Our models are rather generic, requiring no priorknowledge on language, and can hence be applied to matching tasks of differentnature and in different languages. The empirical study on a variety of matchingtasks demonstrates the efficacy of the proposed model on a variety of matchingtasks and its superiority to competitor models.

语义匹配是很多自然语言处理任务的核心。一个成功的匹配算法需要适当地模拟句子的内部结构以及他们的之间的互动关系。作为达成这个目的地重要一步，我们提出用卷积神经网络来模拟两个句子，通过适应视觉和话语的卷积策略。该类模型不仅仅通过一层一层组合和池化很好地表示了句子的层级结构，而且描述了很多不同层级的匹配模式。我们的模型非常具有一般性，不需要语言的先验只是，而且因此可以应用到不同性质的不同语言的匹配中去。实验证明在很多匹配任务中我们提出的方法的效能以及它比其他较强的模型的优越性。

这个英文看起来是中国人写的，非常不地道。不过讲的东西基本说清楚了，但是在问题的提出方面，并没有指出匹配的核心问题在哪儿。