意图理解是新一代电力智能交互平台中的一项基础技术。通过将用户诉求自动分类与分级,可以大幅提升服务效率和质量。针对电力交互平台中的意图理解问题,提出一种基于深度学习的多任务集成模型,该模型可以同时训练意图理解中密切相关的两项子任务：意图检测(Intent Detection)与语义槽填充(Slot Filling)。使用具有长短期记忆(Long-Short Term Memory,LSTM)结构和门控循环单元(Gated Recurrent Unit,GRU)的深度双向循环神经网络(recurrent neural network,RNN)作为基本分类器,多层感知机(Multi-Layer Perceptron,MLP)框架用于组合输出结果,并基于词向量特征与词性特征对模型进行增强。在真实数据上的实验表明该集成多任务模型相比单一模型或其他主流方法更为有效。
英文摘要:
Intent understanding is a fundamental technology in new-type power intelligent interaction platform. Through classifying and grading intents of customers automatically, efficiency and quality of power service can be remarkably improved. Towards intent understanding problem in power interaction platform, an ensemble multi-task model based on deep learning is proposed, which can simultaneously train two closely related sub-tasks in intent understanding: intent detection and slot filling. Recurrent neural networks with long-short term memory (LSTM) and gated recurrent unit (GRU) respectively are used as base model in the proposed model, and multi-layer perceptron generates the final output. Word vectors and part-of-speech (POS) features are used to reinforce the proposed method. Experimental results on real-world dialogue data indicates the superiority of the proposed ensemble multi-task model compared with independent models and other peer models.