一种会话理解模型的问题生成方法

Conversational Comprehension Model for Question Generation

会话问题生成(Conversational Question Generation,CQG)不同于根据段落和答案生成单轮问题的问题生成任务,CQG额外考虑由历史问答对构成的会话信息,生成的问题承接会话历史内容,保持较高的一致性。针对这一特性,文中提出了字级别和句级别注意力机制模块来增强对会话历史信息的提取能力,确保当前轮次的问题融合会话历史中每个词和句子的特征,从而生成连贯的、高质量的问题。疑问词的正确性较重要,生成的问题需要和数据集中原始问题对应的答案类型相互匹配,在疑问词预测模块中构造额外的损失函数作为疑问词类型的限制。综合各个模块得到会话理解模型(Conversational Comprehension Network,CCNet),实验结果表明,该模型在大部分评测指标上高于基线模型,在CoQA数据集上Bleu1和Bleu2分别达到39.70和23.76,生成的问题质量更高。在消融实验和跨数据集实验中该模型被证明是有效的,说明CCNet模型具有较强的通用能力。

Conversational question generation(CQG)is different from the question generation task of generating single-round questions based on paragraphs and answers.CQG additionally considers the conversational information composed of historical question and answer pairs,and the generated questions inherit the historical content of the conversation and maintain high consistency.In response to this feature,the article proposes word-level and sentence-level attention mechanism modules to enhance the ability to extract conversation history information,ensuring that the current round of questions integrates the characteristics of each word and sentence in the conversation history,thereby generating a coherent,high-quality question.The accuracy of the question word is more important.The generated question needs to match the answer type corresponding to the original question in the data set.An additional loss function is constructed in the question word prediction module as a limitation of the question word type.The conversational comprehension network(CCNet)model is obtained by synthesizing each module.Experiments show that this model is higher than the baseline model in most evaluation indicators.On the CoQA dataset,Bleu1and Bleu2reach 39.70and 23.76,respectively,and the quality of the generated questions is higher.The model is proved to be effective in ablation experiments and cross-dataset experiments,indicating that the CCNet model has strong general capabilities.