Chinese Recognizing Textual Entailment Based On Abstract Extraction And Text Enhancement

Recognizing Textual Entailment aims to infer the semantic relationship between two pieces of text: Entailment,Contradiction,Neutral.In this task,it is crucial to enable deep learning models to better understand the semantic meaning of text for the classification of semantic relationships between texts.Currently,most Recognition Textual Entailment methods use the method of mutual attention to determine the semantic relationship between sentences,which can only capture the interaction information between sentences,weaken the global information of sentences,and do not consider the syntactic structure information of sentences.Moreover,these models perform poorly when dealing with low-frequency words.Based on the above problems,this article proposes the following solutions.(1)To address the issue that most deep learning models can only capture the interaction information between sentences and do not consider syntactic structure information,this article proposes a Recognition Textual Entailment model that incorporates syntactic structure and summary information.By combining selfattention and mutual attention mechanisms,this model considers the global and local interaction information of sentences and integrates syntactic structure information to more accurately infer the semantic relationship between sentences.Additionally,a portion of the civil service exam multiple-choice questions was collected and organized,and a summary extraction method was used to solve the problem of length asymmetry caused by lengthy questions and brief answers.Finally,this model and the Recognition Textual Entailment idea were applied to exam question answering.Experimental results demonstrate that the performance of this model on both public datasets and civil service exam questions outperforms multiple benchmark models.(2)To address the issue of poor performance of deep learning models when faced with low-frequency words,this article presents research on Recognition Textual Entailment based on text enhancement.This method divides different low-frequency word sequences based on a frequency threshold and enhances the semantic information of low-frequency words through Sememe information enhancement and synonym replacement.If Sememe or synonyms do not exist,character-level information enhancement is performed.Experimental comparisons reveal that both text enhancement strategies can bring varying degrees of performance improvement,particularly when extracting sentence pairs containing low-frequency words.