动态编码驱动型会话问答方法研究

Research on Dynamic Coding-driven Conversation Question Answering Method

会话问答即多轮问答任务,是对话式人工智能的重要组成部分.如何对复杂的上下文信息进行高效特征提取,一直是会话问答任务的重大难题.现有的方法通常会对其经过多层LSTM处理,很容易产生冗余信息造成上下文偏差.为此,提出动态编码网络的方法,该方法基于Encoder-Decoder框架,但在编码过程融入了动态的方式,以更好地理解段落中的内容和会话历史信息.在动态编码层,编码机制迭代地读取对话历史信息,并且每次迭代的输出都将通过决策器P_(d)与之前的编码表示动态结合,舍弃不相关的信息,生成新的编码表示,随后被送往下一迭代程序中.最终,使模型预测答案的质量更高,整个对话更加流畅连贯.在最新发布的CoQA数据集的实验结果与各种基准和模型变体相比,证明了提出的方法是有效的.

Conversational question answering is a multi-round question answering task,which is an important part of conversational artificial intelligence.How to efficiently extract features from complex context information has always been a major problem in conversational question and answer tasks.Existing methods usually process it through multi-layer LSTM,which can easily generate redundant information and cause context deviation.To this end,this paper designs a dynamic encoding network method,which is based on the Encoder-Decoder framework,but incorporates a dynamic way into the encoding process to better understand the content of the paragraph and the conversation history information.In the dynamic coding layer,the coding mechanism iteratively reads the dialogue history information,and the output of each iteration will be dynamically combined with the previous coding representation through the decision maker P d,discarding irrelevant information,generating a new coding representation,and then sending it Go to the next iteration of the program.In the end,the quality of the model′s prediction answers is higher,and the entire dialogue is more fluid and coherent.The experimental results on the newly released CoQA data set are compared with various benchmarks and model variants,which prove that the designed method is effective.