Cluster-Guided Label Generation in Extreme Multi-Label Classification

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1670–1685 May 2-6, 2023 ©2023 Association for Computational Linguistics Cluster-Guided Label Generation in Extreme Multi-Label Classification Taehee Jung ♣ ∗ Joo-Kyung Kim ♣ Sungjin Lee ♣ Dongyeop Kang ♡ ♣Amazon Alexa AI ♡University of Minnesota {jungtaeh,jookyk,sungjinl}@amazon.com dongyeop@umn.edu Abstract For extreme multi-label classification (XMC), existing classification-based models poorly per- form for tail labels and often ignore the se- mantic relations among labels, like treating “Wikipedia” and “Wiki” as independent and separate labels. In this paper, we cast XMC as a generation task (XLGen), where we benefit from pre-trained text-to-text models. However, generating labels from the extremely large la- bel space is challenging without any constraints or guidance. We, therefore, propose to guide label generation using label cluster informa- tion to hierarchically generate lower-level la- bels. We also find that frequency-based label ordering and using decoding ensemble meth- ods are critical factors for the improvements in XLGen. XLGen with cluster guidance signifi- cantly outperforms the classification and gener- ation baselines on tail labels, and also generally improves the overall performance in four popu- lar XMC benchmarks. In human evaluation, we also find XLGen generates unseen but plausible labels. Our code is now available at https:// github.com/alexa/xlgen-eacl-2023. 1 Introduction Extreme multi-label classification (XMC) is a task to predict multiple relevant labels for a given input where the label space is extremely large. Conven- tional approaches for XMC decompose the prob- lem into a set of binary classifications, training one-vs-all classifiers for each label. However, they encounter several issues in practical use cases. First, the labels in XMC are long-tail distributed. In other words, only a few labels have sufficient positive samples, thereby the other infrequent la- bels could be rarely predicted during inference as we see the heavily right-skewed distribution in the long-tail in Figure 2a. Second, multi-label clas- sification techniques such as one-by-one and la- bel powerset (Gibaja, 2015) assume independent ∗ Part of this work was done during an internship at Ama- zon Alexa AI. wikipedia wiki science interesting fun video funny food humor weird humour wtf #aftedarkclub soda eruption wikipedia fun science diet wiki funny coke tv video health interesting humor food Diet Coke and Mentos Eruption Diet Coke and Mentos Eruption is a reaction of Diet Coke and mint Mentos candies, a bottle of Diet Coke (other carbonated beverages may be used instead) and dropping some Mentos. This causes the Coke to foam at a rapid rate and spew into the air. beverage chemistry coke dietcoke eruption video explosion food fun funny interesting mint science AttnXML (You et al., 2019) Human Labels XLGen (ours) wikipedia wiki research article science experiment mentos diet_coke eruption geyser physical_reaction internet videos GPT-3 (Brown et al., 2020) Figure 1: The predicted and generated labels from At- tentionXML (You et al., 2019), GPT-3 (Brown et al., 2020), and XLGen-BCL, respectively, for Wikipedia page on diet coke and mentos eruption. We marked labels to be correct (blue), wrong (strikethrough), and …

Content truncated to 500 words