A training method, training device and recognition device for intent recognition

Abstract

A training method, device, and recognition device for intent recognition, the method comprising: obtaining a corpus text vector corresponding to corpus text in the corpus text; constructing a joint loss function formula for a training model; obtaining training data; and performing segmentation processing on the training data Map to the corresponding corpus text vector, which is recorded as the training vector; use the training model to predict the training sample vector, and calculate the loss function value of each training model based on the prediction result; calculate the joint loss function value based on the loss function value of each model, and judge Whether the joint loss function value is less than the set threshold, if yes, the training ends, if not, update each model parameter, and continue iterative training. In this way, the generalization ability of intent recognition can be improved as much as possible, and the problems of semantic ambiguity and fault tolerance can be solved.

Description

technical field [0001] The invention relates to the field of intelligent technology, in particular to a training method, training device and recognition device for intent recognition based on multi-task learning. Background technique [0002] "Intent recognition" refers to determining the intent category of a piece of information input by a user to express a query requirement. The current intention recognition technology is mainly used in search engines, human-computer dialogue systems, etc., and can be divided into template / thesaurus-based and supervised classification-based methods. The template / thesaurus-based method mines specific intentions from user historical input. Template / thesaurus, if the user input matches the words in the template / thesaurus of the corresponding category, the input is considered to belong to the intent category; the method based on supervised classification builds an intent classification model based on historical input data to predict user input...

AI Technical Summary

Problems solved by technology

[0003] ①. Generalization ability. Template-based methods are limited by template and lexicon coverage issues. Supervised classification-based methods are limited by training corpus data scale and data quality issues

[0004] ②. Problems of ambiguity, lack of semantics, and fault tolerance. Short texts often have problems such as incompleteness, lack of semantics, and input errors. For example, the user enters "Night Queen of the Playing Bear" and actually wants to search for "Midnight Harem of Teddy Bear" "

However, there are two problems with this method: first, the method obtains text vectors based on other tasks, and then splices them with the text vectors of the target task, and then trains the classifier of this task, and the errors of other tasks may affect this task. Negative impact; secondly, if other tasks are irrelevant to the current task, a large amount of irrelevant external information is introduced, which may cause disturbance to the classification results

Images