提出了一种融合专业知识的意图抽取联合模型.该方法融合了垂直领域的知识图谱,可以在理解用户原本意图的基础上,识别出隐含的其他信息;可以对意图识别与槽填充之间的关系进行显式建模,将二者联合训练,提高模型的整体性能.在公共数据集——航空公司旅行信息系统(Airline Travel Information System,ATIS)及个人语音助手数据集(The Snips Voice Platform,SNIPS)上进行实验,发现该模型结果优于多种意图识别模型.在构建的工业领域质量溯源(quality traceability questions,QTQ)数据集与生产维修(product repair,PR)数据集上进行实验,发现该模型取得了最优的效果,其中在QTQ数据集上的意图识别准确率为87.2％,槽填充F1值为86.7％,全句准确率为75.6％;在PR数据集上的意图识别准确率为92.5％,槽填充F1值为90.1％,全句准确率为88.5％,与其他主流模型相比均有不同程度的提升.实验证明,该模型在工业领域下的意图识别任务具有较高的准确率及泛化能力.
Intent-extraction joint model with expertise integration
This paper proposed an intent-extraction joint model by fusing various ex-pertise,which integrated knowledge graphs of vertical domains so that it could identify hidden information by understanding the user's original intention.The relationship be-tween intent detection and slot filling was explicitly modeled,and the two were jointly trained to improve the overall performance of the model.The proposed model was eval-uated on public datasets,Airline Travel Information System(ATIS)and The Snips Voice Platform(SNIPS),and the results yielded were superior to those of other intent detection models.Additionally,experiments were conducted on the constructed industrial quality traceability questions(QTQ)and product repair dataset(PR),where the proposed model yielded the best results.Specifically,the accuracy of intent detection on the QTQ dataset was 87.2％,the F1 value of slot filling was 86.7％,and the entire-sentence accuracy was 75.6％.Meanwhile,the accuracy of intent detection on the PR dataset was 92.5％,the F1 value of slot filling was 90.1％,and the entire-sentence accuracy was 88.5％,which indicated improvements to varying degrees compared with other mainstream models.Experiments showed that the proposed model could perform intent detection tasks with high accuracy and generalizability in industrial fields.
knowledge graphintent detectionslot fillingindustrial field
万方数据
维普
