Research On Open Domain Question Answering Technology Based On Deep Neural Network And Weakly Supervised Learning

In recent years,artificial intelligence technologies represented by machine learning and deep learning have advanced by leaps and bounds.These technologies not only change our lives,but also make all kinds of information grow explosively.How to quickly and accurately find the required answers from a large amount of information has become a widespread concern of relevant researchers,and Open Domain Question and Answer System came into being under this background.At present,the open domain question answering technology basically adopts the retrieval reading comprehension interactive framework,but the open domain question answering system based on this architecture still has many defects.For example,for reading comprehension models,the answer is usually provided by predicting where the answer in the article begins and ends,which may make the answer provided by the model less concise;For retrieving models,a large amount of supervisory data is usually required to train,and noise is unavoidable in the process of marking positive and negative samples.This paper will improve the reading comprehension model and retrieval model to solve the problems of open domain question and answer technology.In the aspect of reading comprehension model,this thesis puts forward a reading comprehension algorithm based on multitask learning pre-training model.First,by learning the pre-training model on different and highly correlated tasks,the model can capture the commonality of multiple tasks and better learn some features that are difficult to capture in a single task.And then,a generative reading comprehension model is established through the pre-training model,so that the model can not only achieve better results in less training time,but also make the generated answers more concise and clear.In the aspect of retrieval model,this thesis makes a study on the lightweight retrieval model of cross-attention weak supervised learning based on reading comprehension model.First,the cross-attention of the reading comprehension model is used to train the retrieval model so that the training of the retrieval model no longer requires labeled query statements and text.After that,this paper establishes a lightweight retrieval model based on this retrieval model,which greatly reduces the amount of retrieval model parameters and training time while guaranteeing certain retrieval accuracy.