TaxoClass: Hierarchical Multi-Label Text Classification Using Only Class Names

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4239–4249 June 6–11, 2021. ©2021 Association for Computational Linguistics 4239 TaxoClass: Hierarchical Multi-Label Text Classification Using Only Class Names Jiaming Shen♦, Wenda Qiu♦, Yu Meng♦, Jingbo Shang∇, Xiang Ren△, Jiawei Han♦ ♦University of Illinois at Urbana-Champaign, IL, USA, ∇University of California, San Diego, CA, USA △University of Southern California, CA, USA ♦{js2, qiuwenda, yumeng5, hanj}@illinois.edu ∇jshang@ucsd.edu △xiangren@ucs.edu Abstract Hierarchical multi-label text classification (HMTC) aims to tag each document with a set of classes from a class hierarchy. Most ex- isting HMTC methods train classifiers using massive human-labeled documents, which are often too costly to obtain in real-world appli- cations. In this paper, we explore to conduct HMTC based on only class surface names as supervision signals. We observe that to per- form HMTC, human experts typically first pin- point a few most essential classes for the docu- ment as its “core classes”, and then check core classes’ ancestor classes to ensure the coverage. To mimic human experts, we propose a novel HMTC framework, named TaxoClass. Specifi- cally, TaxoClass (1) calculates document-class similarities using a textual entailment model, (2) identifies a document’s core classes and uti- lizes confident core classes to train a taxonomy- enhanced classifier, and (3) generalizes the classifier via multi-label self-training. Our ex- periments on two challenging datasets show TaxoClass can achieve around 0.71 Example- F1 using only class names, outperforming the best previous method by 25%. 1 Introduction Hierarchical multi-label text classification (HMTC) aims to assign each text document to a set of rel- evant classes from a class taxonomy. As a funda- mental task in NLP, HMTC has many applications such as product categorization (Goumy and Mejri, 2018), semantic indexing (Li et al., 2019), and fine- grained entity typing (Xu and Barbosa, 2018). Most existing methods address HMTC in a super- vised fashion — they first ask humans to provide many labeled documents and then train a text clas- sifier for prediction. Many classifiers have been developed with different deep learning architec- tures such as CNN (Kim, 2014), RNN (You et al., 2019), Attention Network (Huang et al., 2019), and achieved decent performance when trained on mas- sive human-labeled documents. Despite such a Document: When our son was about 4 months old, our doctor said we could give him crafted cereal. We bought this product and put it in his bottle. He loved this stuff! This cereal digests well and didn’t lock up his bowels at all. We highly recommend this cereal. Root diapering baby product baby formula grocery & gourmet food nursery feeding crafted cereal toddler fruit beverages baby food baby cereal ������ ������ ������ ������ ������ Labeled Class Core Class Class Taxonomy Figure 1: An exemplar document tagged with five classes. Here, if we are able to pinpoint this document’s most essential classes, crafted cereal and baby cereal, as core classes, we can check their ancestor classes in the taxonomy and recover all the …

Content truncated to 500 words