Using Fractional Latent Topic to Enhance Recurrent Neural Network in Text Similarity Modeling

Abstract

Recurrent neural networks (RNNs) have been widely used in text similarity modeling for text semantic representation learning. However, referring to the classical topic models, a text contains many different latent topics, and the complete semantic information of the text is described by all the latent topics. Previous RNN based models usually learn the text representation with the separated words in the text instead of topics, which will bring noises and loss hierarchical structure information for text representation. In this paper, we proposed a novel fractional latent topic based RNN (FraLT-RNN) model, which focuses on the text representation in topic-level and largely preserve the whole semantic information of a text. To be specific, we first adopt the fractional calculus to generate latent topics for a text with the hidden states learned by a RNN model. Then, we propose a topic-wise attention gating mechanism and embed it into our model to generate the topic-level attentive vector for each topic. Finally, we reward the topic perspective with the topic-level attention for text representation. Experiments on four benchmark datasets, namely TREC-QA and WikiQA for answer selection, MSRP for paraphrase identification, and MultiNLI for textual entailment, show the great advantages of our proposed model.