Abstract
In recent years, positional information has shown good performance in deep neural networks for text matching. Most positional deep neural networks focus on modeling positional information based on the word-level matching signals, whereas the positional influence and interaction among texts have not been well studied during the generation of matching scores. In this paper, we propose a novel positional convolution neural matching model that holds positional influence and interaction in multiple perspectives for text matching. To be specific, we first encode the perspectives of positional information at the word level, the phrase level, and the sentence level. Then, a position-similarity mapping layer is defined to project word-level positional information to local matching signals, which bridges the gap between the word embedding input layer and the hidden convolutional layer. After that, a position-sensible convolution filter is proposed to capture and extract positional information at the phrase level and the sentence level. In particular, we assume that a phrase or sentence has an influence on its neighboring phrase or sentence, and the position-sensible convolution filter is generated on the basis of influence propagation, instead of a random matrix, as in the traditional convolutional neural network. Finally, we offer a multiple-perspective matching function to aggregate positional information at the word level, phrase level, and sentence level. Three standard datasets are used to evaluate our approach, namely ClueWeb-09-Cat-B for web search and TREC-QA and WikiQA for answer selection. It is notable that we achieve the new state-of-the-art performance on ClueWeb-09-Cat-B. Furthermore, on TREC-QA and WikiQA, our model outperforms all deep neural network approaches without an attention mechanism, and is comparable to if not better than approaches that rely on attention mechanisms.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have