Research And Application Of Reading Comprehension-based Qa Techniques Based On Pre-trained Models

Reading comprehension-based QA technology,known as machine reading comprehension,is the current mainstream technical solution for QA systems and a subtask of natural language processing.The task aims to use computers to answer a given question by reading textual information.It is considered to be of great research value as it can be a measure of how well a machine understands natural language.Along with the rapid development of deep learning technology in recent years,researchers had made great progress in reading comprehension technology but are still faced with many difficulties.For example,the model structures of reading comprehension are relatively homogeneous,which will affect the accuracy of question-answering.Also,the reading comprehension model is only designed for single candidate document scenario,so it is difficult to be applied to open domain QA scenario.To address these challenges,this paper takes extractive reading comprehension and open domain QA as research topics,and design new QA models by combining the mainstream techniques of deep learning such as pre-trained language model and attention mechanism.The main research contributions of this paper are as follows.To address the problem of homogeneous reading comprehension model,this study proposes an attention-based reading comprehension model that emulates the human reading process.The model utilizes a pre-trained language model instead of the recurrent network to improve the encoding efficiency.Moreover,it uses a mutual attention mechanism to capture matching information between questions and passages.Finally,the effectiveness of the model is verified through experimental comparison with current mainstream models.To address the difficulty of application to open domain QA,this study proposes a multi-step pipeline network,which is named as retrieve-rerank-reading network.First of all,the network utilizes a pre-trained language model as retriever to enhance semantic retrieval effect.Then,it makes use of a multi-step model to finish answer selection.The multi-step model consists of a rerank network and a reading comprehension network and is trained in multi-task.The multi-step model shares the encoding network to improve efficiency and achieves information complementation to enhance the reading effect via multi-task training.The experimental results show that the present network has better performance compared with the baseline models.Finally,by synthesizing above research work,this study implements a reading comprehension-based QA system,using retrieve-rerank-reading network as the core strategy of answer response.The QA system is built with Web technology.It assigns logical functions to different layers,achieves data persistence by using vector index libraries and offers a web page to facilitate user's interaction.