A Text Mathcing Model Combining Syntax And Semantics

Text matching plays a key role in natural language processing.Many natural language processing problems,such as information retrieval,machine translation and question answering,are based on text matching.So,the improvement of text matching means a lot to many natural language processing problems.Traditional text matching models are based on artificial features,which makes them have a high dependency on the experience of experts.And the widely applied features are based on the accurate matching of words.Besides,the semantic relationships between words are ignored.So,these models are not able to acquire a deep understanding of texts,which limits their performance in text matching.With the breakthrough of deep learning,many text matching models based on deep learning are issued.Most of them try to extract significant patterns from texts using word embedding as a tool.Accurate text matching requires deep semantic matching.So,combining lexical semantics and dependencies between words is meaningful to semantic matching.To date,both traditional and deep text matching models are unable to fully make use of lexical semantics and dependencies at the same time,which seriously limits their performances on text matching.To perform more accurate text matching,the model issued in this paper first combines word embedding and dependency parsing to generate semantic representations for text.Then a matching matrix is constructed for two pieces of text by cosine mean convolution and K-Max pooling based on the generated semantic representations.Finally,a Long-Short Term Memory(LSTM)is applied to map the matching matrix to a true matching level.Experiments are conducted on a full-text dataset compared with other nine text matching models.The model issued in this paper noticeably outperforms a traditional text matching model and other eight deep text matching models.So,considering lexical semantics and dependencies as a whole is significant to text matching.