Question and answer reading comprehension model acquisition method, device, electronic device and storage medium

Abstract

This application discloses a question and answer reading comprehension model acquisition method, device, electronic equipment, and storage medium, involving the field of deep learning, wherein the method may include: for N models with different structures, N is a positive integer greater than one, respectively using unsupervised training The data is pre-trained to obtain N pre-training models, and different models correspond to different pre-training tasks; the main task is the question-and-answer reading comprehension task, and other predetermined natural language processing tasks are used as auxiliary tasks, using supervised training data Fine-tune each pre-trained model to obtain N fine-tuned models; determine the final required Q&A reading comprehension model based on the N fine-tuned models. Applying the solution described in this application can improve the generalization ability of the model, etc.

Description

technical field [0001] This application relates to computer application technology, in particular to a method, device, electronic device and storage medium for acquiring a question-and-answer reading comprehension model in the field of deep learning. Background technique [0002] Question and answer reading comprehension technology refers to given one or more paragraphs (P) and a question (Q), using machine learning methods to let the model predict the answer (A). [0003] The current question-and-answer reading comprehension models are mostly obtained based on pre-training-fine-tuning methods, that is, first select a model structure, then pre-train on a large number of unsupervised training data from a single source, and then use supervised training data, in a single The Q&A reading comprehension task is fine-tuned to obtain the final required Q&A reading comprehension model. [0004] However, the model structure and training tasks of the above method are single, which mak...

AI Technical Summary

Problems solved by technology

[0004] However, the model structure and training tasks of the above method are single, which makes it difficult for the model to learn some common features, which leads to weak generalization ability of the model, etc.

Images