aMV-LSTM: an attention-based model with multiple positional text matching

Abstract

Deep models are getting a wide interest in recent NLP and IR state-of-the-art. Among the proposed models, position-based models and attention-based models take into account the word position in the text, in the former, and the importance of a word among other words in the latter. The positional information are some of the important features that help text representation learning. However, the importance of a given word among others in a given text, which is an important aspect in text matching, is not considered in positional features. In this paper, we propose a model that combines position-based representation learning approach with the attention-based weighting process. The latter learns an importance coefficient for each word of the input text. We propose an extension of a position-based model MV-LSTM with an attention layer, allowing a parameterizable architecture. We believe that when the model is aware of both word position and importance, the learned representations will get more relevant features for the matching process. Our model, namely aMV-LSTM, learns the attention based coefficients to weight words of the different input sentences, before computing their position-based representations. Experimental results, in question/answer matching and question pairs identification tasks, show that the proposed model outperforms the MV-LSTM baseline and several state-of-the-art models.