Hierarchical Label Generation for Text Classification

Findings of the Association for Computational Linguistics: EACL 2023, pages 625–632 May 2-6, 2023 ©2023 Association for Computational Linguistics Hierarchical Label Generation for Text Classification Jingun Kwon1,3, Hidetaka Kamigaito1,2, Young-In Song3, and Manabu Okumura1 1Tokyo Institute of Technology 2Nara Institute of Science and Technology (NAIST) 3Naver Corporation jingun.kwon@navercorp.com kamigaito.h@is.naist.jp song.youngin@navercorp.com oku@pi.titech.ac.jp Abstract Hierarchical text classification (HTC) aims to assign the most relevant labels with the hier- archical structure to an input text. However, handling unseen labels with considering a label hierarchy is still an open problem for real-world applications because traditional HTC models employ a pre-defined label set. To deal with this problem, we propose a generation-based classifier that leverages a Seq2Seq framework to capture a label hierarchy and unseen labels explicitly. Because of no available social media datasets that target at HTC, we constructed a new (Blog) dataset using pairs of social media posts and their hierarchical topic labels. Exper- imental results on the Blog dataset showed the effectiveness of our generation-based classifier over state-of-the-art baseline models. Human evaluation results showed that the quality of generated unseen labels outperforms even the gold labels. 1 Introduction Hierarchical text classification (HTC) aims to as- sign the most relevant labels with their structure for a given document. Because real-world applications categorize documents into a structured class hier- archy sequence (Silla and Freitas, 2011), such as patent collections (Tikk et al., 2005), web content collections (Dumais and Chen, 2000), and medical record coding (Cao et al., 2020), it is needed to cap- ture the label hierarchy for better categorization. To solve the HTC task, recent work has focused on enhancing label embeddings with a taxonomic hierarchy (Cao et al., 2020; Zhou et al., 2020; Wang et al., 2021) or considering a sequential classifica- tion approach (Rivas Rojas et al., 2020; Yang et al., 2018, 2019) that leverages a Seq2Seq framework to capture the label hierarchy. Despite the previous methods being successful, their approaches classify labels sequentially by choosing them from the pre- defined label set in the training dataset. It is still an open problem for real-world applications to handle Seq2Set Seq2Gen Powdered milk Parenting, Marriage Topic Sub-Topic Input Text Seq2Seq Model Pre-defined Label Set Seq2Seq Model ‘Parent', ‘ing’, ‘Marriage’, ‘>’, ‘Pow’, ‘d’, ‘ered’, ‘milk Input Text Shared Dictionary Sub-word sequence : Pre-defined : Text Figure 1: Different from previous Seq2Set (Rivas Rojas et al., 2020), our Seq2Gen can handle unseen labels with sub-word level generation. unseen labels that do not appear in the pre-defined label set from the training dataset (Banerjee et al., 2019; Aly et al., 2019; Xu et al., 2021). Due to se- vere deficiencies in annotating data for labels in a hierarchy and handling unseen labels for real-world applications (Liu et al., 2021), we need a general modeling framework for handling unseen labels while explicitly incorporating a label hierarchy to overcome the restriction of the pre-defined label set for the development of real-world text classification applications. For this purpose, we propose a generation-based classifier that can generate unseen labels …

Content truncated to 500 words