Sentence Pair Semantic Enhanced Matching Network for Text Information Retrieval.

Text semantic matching is a core problem in Natural Language Processing (NLP), such as information retrieval and question answering, which is significant in intelligent human-computer interaction. However, most deep neural matching models are driven by external knowledge, which lacks fine-grained feature extraction from the sentences, leading to limited performance improvement. Therefore, this paper focuses on generating sentence semantic representations without external knowledge for sentence pair matching. We propose a Gated Attentive Convolutional Recurrent Neural Network (GACRNN), which incorporates a Gated Convolutional Neural Network (GCNN), Multi-scale Cross-Channel Attention Block (MC2AB), and bidirectional gate recurrent units (BiGRUs). First, a gate mechanism is introduced in the convolutional neural network to control the information interaction to extract multi-scale features from the sentence. Then, a multi-scale cross-channel attention mechanism is utilized to capture the feature dependencies at different scales in the channel dimension to generate expressive sentence representation. Finally, an extensive evaluation is conducted on two open-domain and two restricted-domain datasets. The experiment results show that the proposed model outperforms other baselines in terms of sentence pair semantic matching accuracy.