Deep Hierarchical Attention Networks for Text Matching in Information Retrieval

Text Matching is the task of examining two pieces of texts, such as query and documents, and determining whether they have the same meaning. Text Matching is very important in many NLP tasks, such as document retrieval, question answering, automatic conversation, machine translation, etc. In recent years, there existed some representation-based and interaction-based neural networks which have achieved some improvements. However powerful attention mechanism is rarely used in these models. Inspired by the success of attention in machine translation and document classification, in this paper, we propose a Deep Hierarchical Attention Networks for Text Matching, namely Deep-HAN-Matching. Specifically, Deep-HAN-Matching extracts meaningful matching patterns and rich contextual features hierarchically from words to total document at the query term level using the recurrent neural network and attention mechanism, and finally rank the matching score produced by the fully connected neural network. Experimental results on WikiQA, a popular benchmark dataset for answer sentence selection in question answering, show that our model can significantly outperform traditional retrieval baseline models and some recent deep neural network based matching models.