Problem generation method based on progressive multi-discriminator

Abstract

The present invention relates to the technical field of problem generation, and more particularly, to a problem generation method based on a progressive multi-discriminator. As use herein, a generatoris used to generate the problem, a discriminator is used to evaluate the problem, and three kinds of discriminators are designed in this paper, namely, a true / false discriminator is used to judge whether the problem is smooth and reasonable, an attribute discriminator is used to judge whether the problem belongs to the category corresponding to the answer, and a question-answer discriminator is used to judge whether the problem can be answered by the corresponding answer. The invention aims at the question-answer mismatch problem in the text generation task, the article adds the answer attribute information to the encoder and decoder in the generator, and a progressive multi-discriminator is designed to strengthen the degree of constraint from easy to difficult, first to ensure the semantic quality of the generated questions, then to constrain the types of questions, and finally to constrain the direct answers of the questions, and to strengthen the matching degree between the questions and the answers.

Description

technical field [0001] The present invention relates to the technical field of question generation, and more specifically, to a question generation method based on progressive multi-discriminators. Background technique [0002] This task belongs to a text generation task, which generates a corresponding question for the article and the specified answer, so that the question can be answered with the answer in the original text. It can be used in consultation systems, counseling systems, fairy tale questions, factual question and answer data generation, etc. It can also be used as a means of data reprocessing to expand the data set for question answering tasks. The question-and-answer data set can be used for question generation during training. In actual use, named entity recognition is performed on articles to extract pairs of entities, which can be used as answers for questions. [0003] The traditional approach is to extract key entities through rules based on syntax tre...

AI Technical Summary

Problems solved by technology

[0006] In the existing research, the fluency and semantic rationality of generated questions can achieve a relatively ideal effect, but there is still a lot of room for improvement in the matching degree of answers. At present, the main method is to encode and learn the answers as answer constraints. Adding it to the output of Decoder to predict the distribution of generated words, adding answer constraints on the basis of encoder-decoder can indeed greatly improve the matching degree of answers, but this constraint is not strong enough to completely solve the problem of mismatched answers, and further constraints need to be strengthened

[0007] In the study of generative confrontation, if a binary classifier is used as the discriminator, then the discriminator will be relatively simple, relatively easy to train, and its accuracy will usually exceed that of the generator, and it is difficult to coordinate between the generator and the discriminator; if the question-answering model As a discriminator, the discriminator will be more complicated, and it is not easy to adjust the model

Images