Sentence Representation And Reasoning Based On Semantics

With the development of artificial intelligence,people are increasingly hoping that computers can understand human language through natural language technology,learn to think like human beings,and finally replace humans to complete difficult tasks with cognitive ability.Sentence representation reasoning technology is the key technology of natural language understanding.The current research content mainly focuses on two aspects: sentence representation method and reasoning model.Although the performance is improved,there are still incomplete semantic expression of sentences,reasoning model lack of inference depth,the lack of interpretability in the reasoning process.In view of the above problems,this thesis studies the sentence representation reasoning technology,the specific research work is as follows:(1)Aiming at the problem that the sentence representation method is not comprehensive,this thesis designs a multi-layer semantic representation network which is applied to sentence representation,and the semantic information of different levels of the sentence is obtained through the multi-attention mechanism.At the same time,by adding the relative position mask between words,incorporate the word order information of the sentence to reduce the uncertainty brought by the word order.Finally,the network is validated on the task of text recognition and emotion classification.The experimental results show that the multi-layer semantic representation network can promote the accuracy and comprehensiveness of sentence representation.(2)Aiming at the lack of inference depth and interpretability of the inference model,a semantic fusion deep matching network is designed,which mainly includes coding layer,matching layer,dependent convolution layer,information aggregation layer and inference prediction layer.Based on the deep matching network,the matching layer is improved,the heuristic matching algorithm is used to replace the two-way long and short memory neural network to simplify the interaction fusion,which improves the inference depth and reduces the complexity of the model.The dependent convolutional layer uses the tree convolutional network to extract the structural information of the sentence along the dependency tree structure of the sentence,which improves the interpretability of the reasoning process.Finally,the performance of the model is verified in several data sets.The verification results show that the reasoning effect is better than the shallow reasoning model,and the accuracy rate on the SNLI test set reaches 89.0%.At the same time,the semantic correlation analysis results show that the dependent convolutional layer can obviously help to improve the interpretability of the reasoning process.(3)The improved optimization of the existing sentence representation reasoning method only considers the limitations brought by one aspect of the sentence representation module or the inference model.This thesis combines the multi-layer semantic representation network with the semantic fusion depth matching network,and proposes a joint optimization method based on multi-layer semantics and SCF-DMN to explore the influence of sentence representation and reasoning model on reasoning performance.The method is implemented in the field of text-based implication identification.The experimental results show that the performance of the joint optimization characterization method is better than the existing method,but the results is between the improved sentence characterization module and the improved method of the separate optimization reasoning model.It indicates the sentence characterization and the reasoning model can promote the reasoning performance,and the optimization of the inference model has a greater impact on the final reasoning result.Comparing the performance changes of each module under joint optimization,it is found that there is a mutual constraint between the sentence representation module and the inference model,which restricts the optimal performance of each module,resulting in no linear superposition of the inference performance after joint optimization.