Research And Implementation Of Reading Comprehension Technology Based On Hybrid Reasoning

In recent years,with the development of machine learning technology,more and more challenging machine reading comprehension data sets have been proposed,and machine reading comprehension technology has become a hot research direction.Recently,two reading comprehension datasets,Re Clor and Logi QA,which test the logical reasoning ability of machines,have attracted the attention of researchers.At present,the mainstream pre-trained language models have been proved not to have good logical reasoning ability because they only focus on word level semantics and do not have good ability to capture text logical relations.Combined with the advantages of pre-trained models and symbolic models,this paper proposes a reading comprehension model based on hybrid reasoning.By deepening the network depth of pre-trained model,this paper improves the ability of logical reasoning of the model and the accuracy of Re Clor and Logi QA datasets.The specific research contents of this paper are as follows:1.A reading comprehension symbolic model based on logical reasoning is proposed.Aiming at the problem that neural network models can not capture symbolic logic,the parser and predefined rules are used to extract the explicit logic in the text,generate the set of logical expressions,and obtain the extended logical expressions through logical theorem reasoning,so as to obtain the logical information in the context and options.2.A reading comprehension model based on hybrid reasoning is proposed.Combining symbolic logic and neural representation,a machine reading comprehension model based on hybrid reasoning is constructed to improve the accuracy of the overall model.By adding additional transformer layers before the output linear layer of the pre-trained model,the feature abstraction and reasoning ability of the model are improved.The DT-Fixup initialization strategy effectively shortens the training time and improves the efficiency of the model.Experiments show that the performance of the model can be significantly improved by transforming the logical expression extracted from the symbolic model into extended text and adding it to the pre-trained model.At the same time,the negative samples are constructed by conditional inversion,deletion or negation of the logical expression to enhance the training data,which further improves the performance of the model.The improvement of the pre-trained model is also helpful to improve the effect of the model.3.Design and implement the reading comprehension question answering system.Based on the reading comprehension model based on hybrid reasoning proposed in this paper,combined with the front and rear development technology,a reading comprehension question answering system is realized,so that users can analyze and get reading comprehension answers through the platform.