Machine reading understanding method based on iterative screening and pre-training enhancement

Abstract

The invention discloses a machine reading understanding method based on iterative screening and pre-training enhancement. The method comprises the following steps: constructing an iterative screening network; training an initial screening network by using problems and document pairs of a HotpotQA data set; finding out the document most relevant to the current document from the original data containing the interference document; and an iterative screening process: screening a support document most related to the initial screening network and the document from the current document set according to the document obtained by the initial network and the original problem, and obtaining a two-step support document related to the problem. Constructing a reader network: finely adjusting the SQuAD data by using a pre-training model to obtain a single-step reading understanding model; and carrying out two times of joint training on the multi-step reading understanding data set to obtain answers and support facts about the questions and the documents. According to the method, a better multi-step machine reading understanding effect can be realized without constructing a complex graph neural network and decomposing a multi-step problem.

Description

technical field [0001] The invention relates to a machine reading comprehension method, in particular to a machine reading comprehension method based on iterative screening and pre-training enhancement. Background technique [0002] In recent years, the rapid development of pre-training models based on large-scale unsupervised datasets has greatly improved the effects of many natural language processing tasks, providing sufficient power for the development of natural language processing. At the same time, with the rapid development of the Internet, information is growing at an explosive rate. Electronic information such as encyclopedias, news, life, books, etc. has gradually become an indispensable part of our life. We contact the world, know the world and understand the world from these electronic information. However, due to the rapid growth of information, it has become increasingly difficult to quickly and efficiently obtain the information we want from the complex data...

AI Technical Summary

Problems solved by technology

[0003] Some existing methods are mainly aimed at the problem of single-step reading comprehension, that is, the question can be answered only based on a single document, but these methods are difficult to solve such problems as "the main literary works of the Nobel Prize winners who once taught at Nanjing University". Multi-step reasoning questions that require multiple documents to answer

For multi-step reasoning question answering, the current main method uses a document filter to filter out documents related to the current question, and then trains a reader to find answers and supporting sentences from related documents, but this method ignores the multi-step The reasoning attributes between question documents cannot make full use of the results of single-step reasoning. At the same time, when solving multi-step reasoning problems, most methods directly use pre-trained models for training, and do not perform on simple single-step reading comprehension datasets. Training, so that the performance of the pre-trained model cannot be enhanced in the field of machine reading comprehension

Images