融合多路注意力机制的语句匹配模型

为了增强语句内关键信息和语句间交互信息的表征能力，有效整合匹配特征，提出一种融合多路注意力机制的语句匹配模型。采用双向长短时记忆网络获取文本的编码信息，通过自注意力机制学习语句内部的关键信息;将编码信息和自注意力信息拼接之后，通过多路注意力机制学习语句间的交互信息;结合并压缩多路注意力层之前和之后的信息，通过双向长短时记忆网络进行池化获得最终的语句特征，经过全连接层完成语句匹配。实验结果表明，该模型在SNLI和MultiNLI数据集上进行的自然语言推理任务、在Quora Question Pairs数据集上进行的释义识别任务和在SQuAD数据集上进行的问答语句选择任务中均表现出了更好效果，能够有效提升语句匹配的性能。

Sentence matching model fused with multi-channel attention mechanism

A sentence matching model fused with multi-channel attention mechanism is proposed to enhance the characterization of key information within sentence and interaction information between sentences, and to effectively integrate matching features. The model first uses a bi-directional LSTM to obtain the encoding information of the text, learns the key information within the sentence through the self-attentive mechanism, and then puts the encoding and self-attentive information together to learn the interaction information between sentences through the multi-channel attention mechanism. Combining and compressing the information before and after the multi-channel attention layer, the final sentence features are obtained by bi-directional LSTM and pooling layer. Finally, the sentence matching is completed by full connection layer. The model is compared with other classical deep sentence matching models on SNLI, MultiNLI, Quora Question Pairs and SQuAD datasets. Experimental results show that the model shows better results in natural language inference tasks on SNLI and MultiNLI datasets, paraphrase recognition tasks on Quora Question Pairs dataset and question-answer sentence selection tasks on SQuAD dataset, and can effectively improve the performance of sentence matching.