Text intention recognition method and system based on projection gradient descent and label smoothing

Abstract

The invention discloses a text intention recognition method and system based on projection gradient descent and label smoothing, and relates to the field of natural language processing question-answering systems. The method comprises the following steps: (1) obtaining an initial vector code through an embedded layer; (2) adding disturbance meeting L2 constraints to an embedded layer by using a projection gradient descent algorithm to form an adversarial sample; (3) using a Transformer network to encode the context semantic information; (4) scaling the category of the real intention by using label smoothing; (5) inputting the encoder output features into a classifier, and calculating the cross entropy between the encoder output features and the smoothed labels; (6) optimizing a target function; and (7) after model training is completed, predicting and outputting intention categories. According to the model, in a classification task, sufficient semantic vector coding can be carried out on an input intention; meanwhile, disturbance is added to a text embedding layer to form an adversarial sample, label sliding is conducted on a final classification target, and the robustness and generalization capacity of the model can be remarkably improved.

Description

Technical field [0001] The invention relates to the field of natural language processing question answering systems, in particular to a method and system for text intention classification based on projection gradient descent and label smoothing. Background technique [0002] With a large number of publicly available online question and answer corpora, the question and answer system has attracted the attention of researchers from industry and academia. Question answering systems are usually based on intelligent products that meet the needs of B-end enterprises, which can significantly improve work efficiency and relieve customer service staff. Its greatest hidden value is to automatically accumulate standardized data in actual scenarios, reduce costs and improve efficiency in mining customer service value information, and can also be used for future precision marketing and product upgrades. The typical application of the question and answer system is to ask questions about a cert...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the problem in the field of intent recognition, the existing text classification model lacks a good trade-off between model complexity and model generalization performance. Too few parameters tend to make the semantic coding ability weak, and the accuracy rate is low in the case of many classification categories. ; Too many parameters make the model too complex and difficult to adapt to the situation of limited training samples

Images