Title: Hybrid Attention Based Neural Architecture for Text Semantics Similarity Measurement
Authors: Kaixin Liu, Yong Zhang and Chunxiao Xing
Abstract: Text semantics similarity measurement is a crucial problem in many real world applications, such as text mining, information retrieval and natural language processing. It is a complicated task due to the ambiguity and variability of linguistic expression. Previous studies focus on modeling the representation of a sentence in multiple granularities and then measure the similarity based on the representations. However, above methods cannot make full use of the diverse importance of different parts in a sentence. To address this problem, in this paper we propose a neural architecture with hybrid attention mechanism to highlight the important signals in different granularities within a text. We first utilize a Bi-directional Long Short Term Memory(BiLSTM) network to encode each sentence. Then we apply the hybrid attention mechanism on top of BiLSTM network. To detect the important parts of a sentence, we adopt a self-attention component to generate sentence level representations and then measure their relevance with a neural tensor network. To better utilize the interaction information, we devise an inter-attention component to further consider the influence of one sentence on another when modeling finer granularity interactions. We evaluate our proposed method on the task of paraphrase identification using two real world datasets. Experimental results demonstrate the superiority of this framework.
