Method and device for training machine reading comprehension model based on answer sampling

Abstract

The invention provides a machine reading understanding model training method and device based on answer sampling. The method includes: Specifically, when machine reading understanding model training is carried out, utilizing a maximum likelihood estimation function to train a target machine reading understanding model, obtaining a basic machine reading understanding model , then utilizing a minimum risk training loss function to continue to train the basic machine reading understanding model so as to achieve fine adjustment of the basic machine reading understanding model parameters, and continuing to optimize the model. Due to the fact that the minimum risk training loss function is adopted to optimize the basic model trained through the maximum likelihood estimation function, answers canbe extracted through the trained model more accurately. In addition, when the minimum risk training loss function is used for loss calculation, k-is carried out on the output answer of the model; Andloss calculation is carried out on answers obtained by top dynamic sampling, so that the data processing amount can be reduced.

Description

technical field [0001] The present application relates to the technical field of natural language processing, in particular to a method and device for training a machine reading comprehension model based on answer sampling. Background technique [0002] At present, deep learning has achieved fruitful results in the fields of image recognition and speech recognition. Machine Reading Comprehension (MRC) has become a new hotspot in the field of artificial intelligence research and application. Its main function is to read and understand a given article or Context to automatically give answers to relevant questions. [0003] With the development of machine reading comprehension technology, the task of reading comprehension is also constantly upgrading, from the early "cloze form" to "single document reading comprehension" based on Wikipedia, such as Stanford SQuAD (Stanford Question Answering Dataset, Stanford Question Answering Dataset) as the task of the dataset. And further...

AI Technical Summary

Problems solved by technology

Therefore, the above-mentioned method of using the likelihood function to maximize the probability of the observed training samples as the goal of the training phase will have the problem of inconsistency with the evaluation indicators used in the testing phase.

Images