Learning Class-Transductive Intent Representations for Zero-shot Intent Detection

Abstract

Zero-shot intent detection (ZSID) aims to deal with the continuously emerging intents without annotated training data. However, existing ZSID systems suffer from two limitations: 1) They are not good at modeling the relationship between seen and unseen intents. 2) They cannot effectively recognize unseen intents under the generalized intent detection (GZSID) setting. A critical problem behind these limitations is that the representations of unseen intents cannot be learned in the training stage. To address this problem, we propose a novel framework that utilizes unseen class labels to learn Class-Transductive Intent Representations (CTIR). Specifically, we allow the model to predict unseen intents during training, with the corresponding label names serving as input utterances. On this basis, we introduce a multi-task learning objective, which encourages the model to learn the distinctions among intents, and a similarity scorer, which estimates the connections among intents more accurately. CTIR is easy to implement and can be integrated with existing ZSID and GZSID methods. Experiments on two real-world datasets show that CTIR brings considerable improvement to the baseline systems.

A Simple Meta-learning Paradigm for Zero-shot Intent Classification with Mixture Attention Mechanism

“…Based on the results, we can make the following observations. [28] 0.7396 0.7206 0.4452 0.4245 CDSSM [4] 0.7588 0.7580 0.4308 0.3765 ZSDNN [15] 0.7165 0.7116 0.4615 0.3897 IntentCapsNet [34] 0.7752 0.7750 0.4864 0.4227 ReCapsNet [19] 0.7996 0.7980 0.5418 0.4769 RL Self-training [37] 0.8253 0.8726 0.7124 0.6587 CTIR [26] 0 • In the standard zero-shot intent classification task, our model outperforms other strong baselines on both SNIPS and SMP, which validates the effectiveness of our model in dealing with zero-shot intent classification. In addition, we can also observe that the pre-trained model BERT can improve our model performance effectively.…”

“…We compare the proposed model with the following state-of-theart baselines: DeViSE [9], CMT [28], CDSSM [4], ZSDNN [15], In-tentCapsNet [34], ReCapsNet [19], RL Self-training [37], CTIR [26] and SEG [35]. SEG is a plug-and-play unknown intent detection method, and it integrates with ReCapsNet in the original paper.…”

“…Ye et al [37] attempt to use the reinforced self-training framework to learn data selection strategy automatically. Si et al [26] propose a class-transductive framework which uses a multitask learning objective function to find the inter-intent distinctions and a similarity scorer to associate the inter-intent connections. However, due to the instability in training and the imbalanced classification shift issue, i.e., the model tends to misclassify the unseen test instances into the seen classes, these methods still struggle in generalized zero-shot learning tasks.…”

Liberating Seen Classes: Boosting Few-Shot and Zero-Shot Text Classification via Anchor Generation and Classification Reframing

“…We conduct our experiments across four varied news and review classification datasets: 20News, Amazon, HuffPost, and Reuters, focusing on few-shot tasks in accordance with the experimental configuration outlined in (Chen et al 2022). Additionally, we follow (Si et al 2021) to evaluate our method on intent classification datasets: SNIPS (Coucke et al 2018) and CLINC (Larson et al 2019) on zero-shot tasks. Note that the training and validation classes are only used in baselines.…”

“…It comprises 22,500 in-scope queries. We follow (Si et al 2021) and use a subset containing 9000 queries.…”