Research On Incorporating Knowledge Into Textual Entailment Recognition

With the development of artificial intelligence technology,people have higher and higher requirements for the ability of machine to understand text.Textual entailment recognition is used to evaluate the level of machine to understand entailment relationship between text pairs,which has gradually become a research hotspot in recent years.Textual entailment recognition aims to recognize the entailment relationship between the premise and hypothesis,and judge the semantic correlation by mining the semantic information between sentence pairs,which is an important part of natural language understanding.This task requires the machine to capture the evidence in sentence pairs to achieve the judgment of entailment relationship,which needs to rely on many kinds of knowledge.In recent years,methods based on pre-trained models have made a breakthrough in this task.However,pre-trained models is difficult to learn world knowledge and language knowledge.Therefore,this paper focuses on the textual entailment recognition method based on knowledge introduction.The main research contents and contributions of this paper include the following two aspects:(1)This paper proposes a multi-source dynamic knowledge integration for textual entailment recognition.Most of the existing methods introduce simple,single and static knowledge,which can not meet the needs of complex reasoning,so we design and implement a multi-source dynamic knowledge integration strategy.This strategy introduces a variety of knowledge,which is encoded by the graph neural network extended based on the transformer model,and then achieves the knowledge interaction through the contextualized self-attention mechanism,so that the proportion of multi-source knowledge can be adjusted adaptively according to the context.The experimental results on SNLI,MNLI and Sci Tail show that the strategy achieves improvements on baseline model,which verifies the necessity of introducing multi-source knowledge.Ablation experiments verifies the complementarity between knowledge and the effectiveness of dynamic knowledge integration.(2)This paper proposes a textual entailment recognition method based on knowledge demand estimation.Previous work of introducing knowledge did not consider the needs of knowledge and the training quality of knowledge representation,which will bring noise to the introduction of knowledge.This paper analyzes the introduction of noise,finds that it is necessary to introduce knowledge on demand according to the understanding level of the text of the model and the training level of external knowledge,and puts forward a knowledge demand estimation method based on reconstruction loss and eigenvector centrality.Experimental results on SNLI,MNLI and Sci Tail show that the proposed method can significantly enhance the pre-trained model,and verify the effectiveness of this method.Ablation experiments also confirm the effectiveness of the knowledge demand estimation strategies and the depth-adaptive interaction method.