Reconstructing Capsule Networks for Zero-shot Intent Classification

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 4799–4809, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 4799 Reconstructing Capsule Networks for Zero-shot Intent Classification Han Liu1∗ Xiaotong Zhang1∗ Lu Fan1∗ Xuandi Fu1 Qimai Li1 Xiao-Ming Wu1† Albert Y.S. Lam2 Department of Computing, The Hong Kong Polytechnic University, Hong Kong S.A.R.1 Fano Labs, Hong Kong S.A.R.2 {cshliu,csxtzhang,cslfan}@comp.polyu.edu.hk {csxfu,csqmli,csxmwu}@comp.polyu.edu.hk, albert@fano.ai Abstract Intent classification is an important building block of dialogue systems. With the bur- geoning of conversational AI, existing sys- tems are not capable of handling numerous fast-emerging intents, which motivates zero- shot intent classification. Nevertheless, re- search on this problem is still in the incipi- ent stage and few methods are available. A recently proposed zero-shot intent classifica- tion method, IntentCapsNet, has been shown to achieve state-of-the-art performance. How- ever, it has two unaddressed limitations: (1) it cannot deal with polysemy when extract- ing semantic capsules; (2) it hardly recog- nizes the utterances of unseen intents in the generalized zero-shot intent classification set- ting. To overcome these limitations, we pro- pose to reconstruct capsule networks for zero- shot intent classification. First, we introduce a dimensional attention mechanism to fight a- gainst polysemy. Second, we reconstruct the transformation matrices for unseen intents by utilizing abundant latent information of the la- beled utterances, which significantly improves the model generalization ability. Experimental results on two task-oriented dialogue datasets in different languages show that our proposed method outperforms IntentCapsNet and other strong baselines. 1 Introduction With the advent of conversational AI, task- oriented spoken dialogue systems are becoming ubiquitous, e.g., chatbots deployed on differen- t applications, or modules integrated in the pop- ular virtual personal assistants like Apple Siri or Microsoft Cortana (Chen et al., 2017). To im- prove business effectiveness and user satisfaction, accurately identifying the intents behind user ut- ∗ Equal contribution. † Corresponding author. terances is indispensable. However, it is extreme- ly challenging not only because user queries are sometimes short and expressed diversely, but al- so because it may continuously encounter new or unacquainted intents popped up quickly from var- ious domains. Conventional intent classification methods (Hu et al., 2009; Tur et al., 2012; Xu and Sarikaya, 2013; Ravuri and Stolcke, 2015; Liu and Lane, 2016; Nam et al., 2016) typically train a su- pervised learning model on large amounts of la- beled data, and are not effective in recognizing e- merging unseen intents. Several zero-shot learning approaches attempt- ed to address the challenges for classifying intents whose instances are not present during training. One common idea is to utilize some external re- sources (Ferreira et al., 2015a,b; Yazdani and Hen- derson, 2015; Kumar et al., 2017; Zhang et al., 2019) such as label ontologies or manually defined attributes. However, such external resources are usually unavailable, as they require substantial ex- tra time and expensive human labour to produce. To implement zero-shot intent classification more easily and intelligently, recent works rely more …

Content truncated to 500 words