Text Semantic Understanding Based On Knowledge Enhancement And Multi-granularity Feature Extraction

Natural language is an important carrier for humans to exchange information,and it is also the key to human-computer interaction.Therefore,how to make machines understand human language is one of the hot researches in academia currently.Since deep learning have been applied to text tasks,more and more researches are based on neural networks.After analyzing the structure and principles of many neural networks,this research studies how to understand the semantics of text.Aiming at the problem of insufficient semantic expression of text in feature extraction and the limitation of knowledge expansion,some improvement measures have been proposed and effective improvements have been made.The main research work of this thesis is as follows:The common methods of semantic feature representation are introduced,and the classification and similarity calculation methods based on semantic understanding is analyzed.The steps and mainstream methods of text classification include methods based on machine learning,and some commonly used neural networks based on deep learning are introduced.The statistics-based methods,knowledge-base-based methods and deep learning-based methods included in the calculation of text similarity are analyzed.Comparing the characteristics of each method,the extracted semantic features are explored,which lays the foundation for subsequent research work.In view of the relatively independent local features obtained by the convolutional neural network,the pooling layer retains the significant semantic features and causes the loss of semantic information.By adding a character convolution structure that cancels the pooling layer after embedding layer,the problem of information loss caused by the pooling layer is avoided,and a dual-granularity semantic representation of "character + word" is formed.Next,after the convolution operation,a global attention mechanism is added to combine the local features extracted by the convolution layer and the global features of the text.At the same time,the model pays more attention to the salient semantic information in the text.Experiments show that the model has a richer understanding of semantic features and has achieved better results in text classification.Aiming at the impact of word order on semantics and the problem that the neural network cannot expand knowledge when extracting features,a multi-granularity feature of "character+word+sentence" after knowledge enhancement is formed based on dualgranularity feature extraction.External knowledge sources based on network information are introduced so that the expanded knowledge in the knowledge base enriches the semantics of the original text.An attention mechanism is added to the coding layer to enhance the information interaction between networks,and the siamese network is used for model training.Experimental results show that the model can effectively enrich semantics to enhance semantic understanding and improve the effect of text similarity calculation.