Natural language reasoning method and system based on syntactic alignment

Abstract

The invention provides a natural language reasoning method and system based on syntactic alignment, and the method comprises the steps: giving a plurality of sentence pairs, dividing words to obtain tokens, and carrying out the coding to obtain word vector features; analyzing the dependency relationship of the sentences through a dependency analyzer to obtain the dependency relationship among the words, and enabling the tokens to correspond to the words; establishing a dependency capturing module, and capturing enhanced representation of syntactic features in sentences; constructing an attention mechanism network, and updating structured features of sentences; taking the structured features as a classification basis to obtain a prediction result; deviation is obtained through calculation of a loss function of the model, and the model is corrected. The natural language reasoning method based on syntactic alignment is adopted, different dependency relationships between words are coded based on the graph network, and compared with a traditional method, the importance of the relationships can be distinguished. Therefore, words input into the model can contain more semantic information after being enhanced by the syntactic context features.

Description

technical field [0001] The present invention relates to the technical field of natural language reasoning, in particular to a method and system for natural language reasoning based on syntactic alignment. Background technique [0002] In recent years, research on natural language processing has continued to deepen. As one of the basic branches, natural language reasoning has emerged with many relatively novel and effective methods. Among them, syntactic features are useful information that has been widely verified. Past research has mainly focused on encoding through tree-like networks and graph-like networks. However, these methods often ignore the unique role of syntactic relationship types and treat all syntactic relationships equally. . [0003] It was found that the Chinese patent of the existing patent document CN112667818A discloses a user comment sentiment analysis method and system that integrates GCN and multi-granularity attention, which uses a syntactic dependen...

AI Technical Summary

Problems solved by technology

[0004] In view of the above-mentioned related technologies, the inability of the inventor to effectively utilize the syntactic structure information is a major existing problem

Images