Class name guided out-of-scope intent classification

• Main Authors: GAUTAM, Chandan, PARAMESWARAN, Sethupathy, KANE, Aditya, FANG, Yuan, RAMASAMY, Savitha, SUNDARAM, Suresh, SAHU, Sunil Kumar, LI, Xiaoli
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2024
• Subjects: Out-of-scope intent classification; Dialogue systems; Class label semantics; Out-of-scope intent detection; Artificial Intelligence and Robotics; Computer Sciences
• Online Access: https://ink.library.smu.edu.sg/sis_research/9752; https://ink.library.smu.edu.sg/context/sis_research/article/10752/viewcontent/EMNLP24Findings_SCOOS.pdf
• Institution: Singapore Management University

The paper introduces Semantics of Class Labelbased Unsupervised Out of Scope Intent Detection (SCOOS), a novel method aimed at enhancing out-of-scope (OOS) intent classification in task-oriented dialogue systems. Unlike prior approaches that rely solely on indomain (ID) data features, SCOOS leverages semantic cues embedded in class labels to improve classification accuracy. The method entails forming a compact feature space centered around the semantics of class labels by minimizing losses between ID features and class names. SCOOS achieves this by creating a compact feature space centered around class label semantics, achieved through minimizing losses between in-domain (ID) features and class names. This involves training two spherical variational autoencoders concurrently to learn a shared latent space between ID features and class names, aligning ID feature data based on the corresponding classes in the latent space, and training a classifier for (m + 1)-class classification using only ID samples, where the (m+1)th class represents OOS samples. Extensive evaluation of three datasets demonstrates that SCOOS outperforms existing methods not only for OOS intent detection but also for ID intent classification. Additionally, an ablation study is conducted to analyze the impact of different components of SCOOS, and we also presented the visualization of the latent space representation providing insights into the influence of semantic information from class labels.