| Summary: | 碩士 === 國立臺灣科技大學 === 資訊管理系 === 106 === In recent year, research interest in Intelligent Virtual Assistant (IVA) has soared in the world. However, current IVA is usually limited to specific domain and only handle a single intent per time. However, people’s intents are usually complex and require several different applications to meet. Several studies explored related issues recently, and however, only a few of studies focused on multi-intent processing of Chinese. Therefore, the purpose of this paper is to propose a multi-intent detection framework of IVA based on Chinese Natural Language Processing. In this paper, people’s intents are categorized into two types which are the explicit intent and the implicit intent. Based on spoken dialogue systems, we propose the Explicit Multi- Intent Processing(EMIP) module and the Implicit Multi-Intent Processing(IMIP) module. EMIP is responsible for recognizing multi-intent from the users’ utterance. IMIP predicts the potential intent of the user based on the users’ utterance and explicit multi-intent.
Finally, we evaluate the performance of EMIP and cross-compare different models which is for processing the users’ implicit multi-intent in two scenarios (the users’ explicit multi-intent is related to each other and the users’ explicit multi-intent is unrelated to each other). The result of our pilot experiment shows that the accuracy of EMIP is 88.2% and the models based on IMIP are better than other models. Moreover, IMIP-ANN-based model has better performance when the users’ explicit multi-intent is related to each other. IMIP-Cluster-based model has better performance when the users’ explicit multi-intent is unrelated to each other.
Regarding the theoretical implications of this paper, the framework we proposed allows IVA to simultaneously recognize explicit multi-intent and implicit-intent from the user's Chinese utterance. We solve the existing problem that IVA can only handle single intent which that people usually cannot express their intent in a sentence. In addition, we provide the performance of different implicit multi-intent models in different scenarios. Regarding the practical implications, the framework we proposed can be applied to any industry that uses IVA. They can use the framework in different fields, different scenarios, and even cross-domain scenarios. In addition, the framework does not require multi-intent-labeled training data. Even if there are only single-intent-labeled training data available, the framework we proposed can still work. We reduce the dependency on multi-intent-labeled training data. Through this way, the cost for labeling multi-intent of data can be reduced, such as time cost, manpower cost.
|
|---|
