The Study And Application Of Sentence Pair Modeling Based On External Knowledge And Memory Network

With the development of natural language understanding technology,intelligent systems such as question answering systems and intelligent customer service are becoming more widely used.The essence of intelligent systems is to model natural language sentences,fully understand the semantics of natural text,and then perform corresponding tasks.Sentence Pair Modeling is the basic task of natural language understanding.It aims to model the semantic information of two sentences and determine the semantic relationship between them.Therefore,the research on sentence pair modeling is helpful for downstream natural language understanding tasks,such as natural language inference and paraphrase identification.Most current neural network-based sentence-pair modeling models rely only on training corpora.On the one hand,they cannot effectively learn the external commonsense knowledge required in the process of understanding texts.On the other hand,confronted with complex sentences,models need more rounds of deep information interactive processing instead of simple neural network node calculation.Therefore,this article will explore the use of external commonsense knowledge and memory networks with multiple rounds of interaction into text modeling methods,and then achieve the goal of high-performance sentence pair modeling.Its main work and contributions include:(1)Knowledge Adaptive Neural Network model: This model acquires common sense knowledge in the form of triples related to sentences from an external knowledge base,and then adaptively injects the encoded external commonsense knowledge into the neural network through a knowledge absorption gate.Finally,the model performs semantic inference and relationship prediction through knowledge-based information aggregation and text inference.The highlight of this model is that external knowledge is included in the sentence modeling process to obtain a priori of factual knowledge,and adaptive semantic absorption is used to achieve more accurate semantic inference.(2)Dual Residual Memory Network model: This model first obtains sentence-level and word-level sentence memory information through a variety of attention mechanisms to form a semantically-rich dual-memory structure.Second,the model is based on the residual memory module with multi-turn semantic inference to make final text semantic prediction.The highlight of the model is that the model obtains semantic information about the sentence at different levels of information granularity,and performs multiple rounds of dynamic information interaction based on the rich information,and then accurately captures the keyword information of the sentence to provide subsequent semantic judgments.(3)Automatic text mark system based on sentence pair modeling: This system provides users with reliable automatic text mark services based on the above two sentence pair modeling models.For the text to be evaluated and the standard reference text provided by the user,the system returns the scoring results of the text to be evaluated and visualizes the scoring details to help users better understand the prediction basis of the model.The system is implemented based on Browser / Server architecture,is simple and easy to use,and can provide users with mark services at any time.