Method for identifying discourse components based on graph neural network

Abstract

The invention discloses a discourse component identification method based on a graph neural network. The objective of the invention is to improve the accuracy of discourse component identification and a Macro-F1 value. According to the technical scheme, firstly, a discourse component recognition system combined with a graph neural network is constructed; preparing a Chinese argument data set required for training the discourse component recognition system; and training the discourse component recognition system by adopting a multi-round circulation mode to obtain an optimal network structure weight parameter, and loading the weight parameter obtained by training to the discourse component recognition system to obtain a trained discourse component recognition system. And the trained discourse component recognition system preprocesses an article input by a user and performs discourse component recognition to obtain a recognition result of the article. According to the method, the purpose of improving the discourse component recognition effect by using the graph neural network technology is achieved, and compared with an existing discourse component recognition method, the Acc value and the Macro-F1 value are both improved.

Description

technical field [0001] The present invention relates to the field of text component recognition, in particular to a text component recognition method based on a graph neural network. Background technique [0002] Natural language processing, referred to as NLP (Natural Language Processing), can be divided into two parts for understanding: "natural language" and "natural language processing". Natural language is a language that is different from computer language, which is the most essential feature that distinguishes human beings from other creatures, because among all creatures, only human beings have the ability to speak. Natural language records the information exchange in the course of human development. A simple "hello", whether in Chinese, English or other languages, or in text or voice, is a part of natural language. Natural language processing, generally speaking, is a technology that uses the natural language used by humans to communicate with machines for interact...

AI Technical Summary

Problems solved by technology

Taking each sentence in an article as a node, if a fully connected graph is constructed (in a graph, there is a path connecting any two vertices), then the time-consuming training of the discourse component recognition system with GNN added will be It is greatly increased, and there is a problem that after deepening the number of GNN layers, the recognition effect of discourse components decreases, because no matter what the initial state of the feature matrix is ​​(randomly generated), after multiple convolutions, the features of all nodes in the same connected component will tend to are consistent, resulting in a decline in the recognition effect of the discourse component recognition system that adds GNN

Another difficulty is how to choose a graph neural network. The graph neural network is divided into four categories, namely: graph convolutional network (GCN), graph attention network (GAT), graph generation network (GGN) and graph spatiotemporal network ( GSN), different graph neural networks will also bring different training time and effects, and need to measure different parameter selections

Images