融合背景知识和常识感知的对话生成

Integration of background knowledge and commonsense perception for dialogue generation

基于背景对话的关键问题之一是知识抽取,但由于有些会话的信息量不足,特别是在一些对话信息较少的情况下,选择恰当的知识变得尤为困难,且目前的生成方式缺乏动态选取背景知识的能力。针对这些问题,提出了KIF模型,引入知识增强库和知识向量并提出知识追踪模块和知识情感反馈模块去解决上述问题。该模型通过双重匹配矩阵的方式获得外部知识与背景知识的权重向量并进行知识选择,在每个解码步长内会根据历史会话和外部知识进行会话生成。最后,在Holl-E和WoW数据集上进行实验,实验结果表明KIF模型相比于之前的模型有明显的性能提升。

One of the key issues in background-based conversation is knowledge extraction.However,due to the insufficient information in some conversations,especially when there is less information in certain dialogues,choosing the appropriate knowledge becomes particularly challenging.Furthermore,the current generation methods lack the capability to dynamically select background knowledge.To address these issues,this paper proposed the KIF model,which incorporated a knowledge enhancement library and knowledge vectors and introduced a knowledge tracking module and a knowledge sentiment feedback module to solve the aforementioned problems.The model obtained weight vectors of external knowledge and background know-ledge through a dual matching matrix method and performs knowledge selection.For each decoding step,the model generated conversation based on historical dialogues and external knowledge.Finally,experiments on the Holl-E and WoW datasets show that the KIF model significantly outperforms previous models.