基于深度神经网络的工单采集模型研究

提出了一种面向工单采集特定场景下的多轮对话模型,实现了机器人自动收集工单的功能。首先,该模型利用卷积神经网络(CNN)将用户输入句子转化成向量,利用信息网络来收集客户咨询的关键性信息,然后用双层长短期记忆网络(LSTM)来记忆上下文信息及管理会话流程,最后采用分类器来预测机器人的回复。实验结果表明,该模型与seq2seq模型相比,需要的训练数据较少,工单完成率高,即使中间信息采集存在误差,最终也可以准确地完成工单的采集。

Intelligent work order collection system based on deep neural net work

A multi-term conversation model for the intelligent work order collection is proposed in this paper.The function of automatically collecting work orders by chatbot is realized for the system.First,we CNN is used to transform user input into vectors.Then belief network is used to collect key information of customer consultation.Next step,double-layer LSTM network is used to memorize context information and manage session flow.Finally,classifier is used to predict chatbot reply.The experimental results show that compared with seq2seq models,this scheme achieve better accuracy with less train data,and even if the intermediate information is collected incorrectly,the collection of work orders can be completed accurately.