Research And Implementation Of Conversational Question Answering Based On Contextual Representation

Machine reading comprehension is an important measure of how well a machine understands human language and is a key step toward artificial intelligence.In recent years,machine reading comprehension has had good applications in search engines and intelligent conversations.Conversational Question Answering is a cross domain between machine reading comprehension and multi-turn conversations,where the machine needs to read documents to answer multiple associated questions.This way is more in line with the general way humans acquire information.Therefore conversational question answering has become a hot topic in academia and industry.The Qu AC dataset is more challenging than other datasets because most of the questions do not exist independently,and only by using the conversation history can understand the current question correctly.Moreover,nearly 20% of the questions in the Qu AC dataset cannot be found in the document and need to be further processed In addition,the length of articles in the Qu AC dataset is longer than many datasets,making it more difficult to find answers.Facing the above challenges,this paper proposes the following three models,the specific contents are as follows:(1)To address the problem of historical information processing,this paper proposes a context-aware model based on the flow mechanism,Context Flow,to further select and model historical contextual information.First,a hierarchical history selection mechanism is used to explicitly select historical question and answer information at word level,sentence level and session level.After that,the session flow is modeled and integrated into the reading inference module to deeply integrate the contextual history information before predicting the answer to the question.The experimental results demonstrate that historical information has a significant impact on the effectiveness of the multi-round machine reading comprehension model.(2)For the types of unanswerable questions,based on the contextflow model proposed in the first work,this paper adds a judgment module for unanswerable questions to form the contextflow + verifier model.Through the two modules of rough reading and intensive reading,we can deeply analyze whether the current problem can find the answer in the document.The experimental results show that increasing the prediction of unanswerable questions is helpful to further improve the accuracy of the model.(3)Aiming at the problem of long reading documents,this paper proposes a multi round machine reading comprehension model pretri model based on segment retrieval.We first divide the article into paragraphs and retrieve the K paragraphs that are most likely to contain answers.Then we use the contextflow model proposed in the first work to read the K paragraphs and give answers.Finally,we synthesize the retrieval module and the reading module to get the final answer.Experiments show that breaking the article into paragraphs for reading is helpful to improve the efficiency of the model and the accuracy of prediction.