A Sentence Encoder Based on Multi-dimensional Semantic Enhancement
Natural language processing tasks such as question answering systems,sentiment analysis,and text classification typically rely on sentence level semantic representations,which are typically obtained by sentence encoders.The existing sentence encoders overly rely on sentence pair interaction information and ignore the semantic representation information contained in the sen-tence itself.Therefore,this article proposes a bidirectional LSTM and maximum pooling hierarchical model(MDSE)based on multi-dimensional semantic enhancement.Firstly,the self attention mechanism is used to capture important features within the sentence.Then,character level embedding information and part of speech information are utilized to form multiple fine-grained feature infor-mation channels.Finally,the outputs of all channels are collectively represented as sentence vectors.Experiments have shown that compared to other deep learning methods,the MDSE model has achieved significant performance improvements on three publicly available natural language inference datasets.Meanwhile,downstream task experiments further validate the sentence representation ability of this method.
attention mechanismsentence representationnatural language inferencelong and short-term memory network
万方数据
维普
