A Probabilistic Framework for Discovering New Intents

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3771–3784 July 9-14, 2023 ©2023 Association for Computational Linguistics A Probabilistic Framework for Discovering New Intents Yunhua Zhou∗, Guofeng Quan∗, Xipeng Qiu† School of Computer Science, Fudan University {zhouyh20,xpqiu}@fudan.edu.cn gfquan21@m.fudan.edu.cn Abstract Discovering new intents is of great significance for establishing the Task-Oriented Dialogue System. Most prevailing approaches either can- not transfer prior knowledge inherent in known intents or fall into the dilemma of forgetting prior knowledge in the follow-up. Furthermore, such approaches fail to thoroughly explore the inherent structure of unlabeled data, thereby failing to capture the fundamental characteris- tics that define an intent in general sense. In this paper, starting from the intuition that dis- covering intents should be beneficial for identi- fying known intents, we propose a probabilistic framework for discovering intents where in- tent assignments are treated as latent variables. We adopt the Expectation Maximization frame- work for optimization. Specifically, In the E- step, we conduct intent discovery and explore the intrinsic structure of unlabeled data by the posterior of intent assignments. In the M-step, we alleviate the forgetting of prior knowledge transferred from known intents by optimizing the discrimination of labeled data. Extensive experiments conducted on three challenging real-world datasets demonstrate the generality and effectiveness of the proposed framework and implementation. Codes is publicly avail- able.1 1 Introduction Unknown intent detection (Zhou et al., 2022) in the Task-Oriented Dialogue System (TODS) has gradually attracted more and more attention from researchers. However, detecting unknown intent is only the first step. For the TODS, intent discovery is crucial but also more challenging. Because the pre-defined intent set in the TODS is limited to cover all intents, the TODS should discover poten- tial new intents automatically during interactions ∗Equal contribution. †Corresponding author. 1https://github.com/zyh190507/Probabilistic-discovery- new-intents with the users. And as a practical matter, a large number of valuable unlabeled data will be gener- ated within the interaction between users and the dialogue system. Considering the limited labeled corpus and time-consuming annotating, which also requires expertise, the TODS should adaptively dis- cover intents from those unlabeled data with the aid of limited labeled data. Just as discovering new intents plays a crucial role in establishing the TODS, discovering new in- tents has raised a lot of research interest just like un- known intent detection. Unsupervised cluster learn- ing is a popular paradigm to solve this problem. Specifically, previous works (Hakkani-Tür et al., 2013, 2015; Shi et al., 2018; Padmasundari, 2018) formulate intent discovery as an unsupervised clus- tering process. However, these methods mainly focus on constructing pseudo-supervised signals to guide the clustering process while neglecting the prior knowledge embedded in the available la- beled data. In real user-facing scenarios, we often possess a small amount of labeled data in advance, which contains prior knowledge that can guide the intent discovery process, and a substantial volume of unlabeled data generated in the interaction with the dialogue system mentioned above, which con- tains both known intents …

Content truncated to 500 words