Research On Reading Comprehension Technology For Open Domain Question Answering

This paper mainly studies the application of machine reading comprehension in open domain question answering system,and designs a question and answer system combining retrieval system and machine reading comprehension.We hoped that our work will promote the application of machine reading comprehension.Open domain means that the content of the question is not limited to the domain.Machine reading comprehension is an emerging answer extraction technology in recent years,which locates accurate answers by predicting the start and end positions of answers.We chose Du Reader as the experimental data,because the questions in Du Reader are sorted out from the real user search records of the search engine.Compared with the deliberately marked question data,this kind of question can truly and objectively reflect the needs of the question answering system.And the experimental results can truly reflect the actual application effect.The research content of our work mainly includes the following aspects.First,we design a machine reading comprehension model for multi-passages.The so-called multi-passages refer to more than one candidate passages corresponding to each question.At present,most machine reading comprehension tasks are focusing on a single passage.Multiple passages can significantly improve the recall rate of the answers and improve the error tolerance of the retrieval system.We improve the model of machine reading comprehension for single-passage,making it suitable for multi-passage reading comprehension.Secondly,we find that the model positioning answer passage is of great significance to extract answer.Simply designing the model as a model of positioning the passage first and then extracting the answer is easy to cause error accumulation.In order to let the model has the ability to locate answer passage and extract answer at the same time,we add a passage ranking subtask to the model and using multi-task learning during training.And the experimental results show that this method does improve the model performance.Thirdly,this paper studies the adaptation of the machine reading comprehension model from the open domain to the specific domain.Like many deep learning-based models,the machine reading comprehension model also has the problem of insufficient data in specific fields.In order to solve this problem,we designed two transfer learning schemes.One is the classic transfer learning that use pre-training and fine-tuning without changing the model?The other is transfer learning combined with adversarial learning.The application of adversarial learning is to realize the transfer of knowledge from open domain data to specific domain data by cheating the discriminator to recognize the domain.And it also improves the model performance and reducing the cost of training time.Finally,this paper integrates the above technology into a question and answer system to simulate its effect in practical application scenarios.In order to better evaluate the question answering system,we used a manual evaluation method.In addition,we also analyzed the efficiency of the system and the existing problems to provide guidance for the future direction of work.