Research On Text Entailment Tasks Based On Mixed Models

With the rapid development of the big data industry,the field of artificial intelligence has attracted the attention of the public,and the text data in the form of electronic documents is also increasing day by day.These massive amounts of data contain very rich information.If the computer can learn the knowledge of human language from the massive text data,it can replace human beings to do many complicated and dangerous tasks and reduce the burden on human beings.The task of textual implication is a task of studying the directional semantic connections between texts.It contains three kinds of relations: implication,neutrality,and contradiction.These relations generally exist in natural language texts.The current research methods are mainly based on the two core technologies of semantic representation technology and semantic reasoning technology.This article studies the corresponding improvement schemes from two issues.The specific contributions can be summarized as the following points:(1)Aiming at the problem of insufficient,complex and slow extraction of sentence semantic information in traditional sentence semantic representation models,a method of text implication recognition based on mixed residual features and self-attention mechanism encoding is proposed.The feedforward neural network composes the sentence encoder,uses multiple methods to calculate the similarity features between sentence pairs,and adds enhanced residual connections to the entire model to mix multiple features to improve the performance of the hybrid model.Finally,the validity of the experimental model is verified on two textual implication task data sets.The experimental results prove that the implication recognition model that mixes multiple feature information can improve the accuracy of textual implication relationship recognition to a certain extent.(2)Aiming at the insufficiency of the modeling relationship between sentence pairs in the semantic reasoning layer,an interactive matching network based on the matching aggregation framework is designed.Use a variety of attention functions to construct a multi-channel matching layer,add syntactic structure information based on a syntactic parse tree to the aggregation layer,use a tree-type long and short-term memory network to extract sentence grammatical structure information,and mix multiple attention interaction features and syntax Information constitutes the final global reasoning information.Finally,experiments and analysis are carried out on the textual implication task data set and the semantic similarity data set.The experimental results prove that the multi-path matching layer plays a certain role in promoting the information exchange between sentences.