Research On Text Matching Method Based On Multi-view Interaction

Text matching aims to determine the semantic similarity between two pieces of text.It plays an important role in many applications,e.g.,retrieval-based dialogues,news recommendation,and natural language inference.Existing works can be divided into three categories: representation-based,interaction-based methods,and methods based on pre-trained language model(PLM-based methods).Among them,PLM-based methods require significant computational resources and have long inference time,which limits their applicability.Compared with representation-based methods,interaction-based methods can capture the relevant features between texts to enhance the text representation,achieving better performance.However,the existing interaction-based methods usually only consider modeling the interaction between texts on a single granularity,leading to limited relevant features and model performance.Therefore,this thesis focuses on interactions from multiple perspectives to capture richer relevant features for improving the performance of text matching task.The contributions of this thesis are mainly reflected in the following three aspects:⑴ Existing methods often have the problem of ignoring the relevant features of multi-granularity information between texts.When capturing the relevant features between texts,existing methods usually only consider the relevant features of a single granularity information,while ignoring the relevant features of other granularity information.To capture the relevant features of multi-granularity information between texts,this thesis proposes a text matching method based on the character and word perspectives interaction,which fully utilizes the character and word granularity information of the text.Based on the character and word granularity information of the text,the method captures the relevant features of different granularity information between texts separately.In addition,to extract the deep semantic information of the text,this thesis combines character and word features of the same text to generate the deep semantic information of the text.Experimental results show that this method can effectively capture and utilize the relevant features of multi-granularity information between texts,and thereby improve the performance of text matching model.⑵ Existing methods often have the problem of ignoring the characteristics of language of input texts.When modeling text matching,existing methods do not consider using the characteristics of language to enrich the relevant features of the text,such as the pinyin and radical of Chinese.To fully utilize the characteristics of language,this thesis proposes a text matching method based on the Chinese characteristics perspective interaction,which fully utilizes pinyin and radical information.According to the semantic features of characters,words,pinyin,and radical,a soft alignment attention network is used to capture the relevant features within and between texts.Experimental results show that this method can fully utilize the Chinese characteristics information,thereby capturing richer relevant features and improving the performance of text matching model.⑶ Existing methods often have the problem of insufficiently exploiting relevant features.Existing methods often only use a single attention mechanism to capture relevant features when processing text,while neglecting the use of multiple attention mechanisms to capture the relevant features within and between texts simultaneously.To explore richer relevant features,this thesis proposes a text matching method based on multiple attention perspective interactions,which models rich relevant features within and between texts based on multiple attention mechanisms.Experimental results demonstrate that the proposed method can effectively capture richer relevant features within and between texts,and improve the performance of text matching model based on these features.