Hierarchical Multi-label Text Classification with Horizontal and Vertical Category Correlations

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2459–2468 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 2459 Hierarchical Multi-label Text Classification with Horizontal and Vertical Category Correlations Linli Xu1, Sijie Teng1, Ruoyu Zhao1, Junliang Guo1, Chi Xiao1, Deqiang Jiang2, Bo Ren2 1Anhui Province Key Laboratory of Big Data Analysis and Application, School of Computer Science and Technology, University of Science and Technology of China 2Tencent YouTu Lab linlixu@ustc.edu.cn, {yunmo, zry1997, guojunll, xiaochi}@mail.ustc.edu.cn {dqiangjiang, timren}@tencent.com Abstract Hierarchical multi-label text classifica- tion (HMTC) deals with the challenging task where an instance can be assigned to multiple hierarchically structured categories at the same time. The majority of prior studies either focus on reducing the HMTC task into a flat multi-label problem ignoring the vertical category correlations or exploiting the dependencies across different hierarchical levels without considering the horizontal correlations among categories at the same level, which inevitably leads to fundamental information loss. In this paper, we propose a novel HMTC framework that considers both vertical and horizontal category correlations. Specifically, we first design a loosely coupled graph convolutional neural network as the rep- resentation extractor to obtain representations for words, documents, and, more importantly, level-wise representations for categories, which are not considered in previous works. Then, the learned category representations are adopted to capture the vertical dependencies among levels of category hierarchy and model the horizontal correlations. Finally, based on the document embeddings and category embeddings, we design a hybrid algorithm to predict the categories of the entire hierarchical structure. Extensive experiments conducted on real-world HMTC datasets validate the effectiveness of the proposed framework with significant improvements over the baselines. 1 Introduction As a fundamental problem in natural language pro- cessing (NLP), text classification is the task of as- signing a given document to one or multiple cate- gories according to its textual content. In practice, many documents are tagged with multiple cate- gories that can be organized in a tree or a Directed Acyclic Graph (DAG) (Wehrmann et al., 2018), which poses a more challenging task. These cat- egories can be organized into different levels of the hierarchical structure, and the task of assigning multiple hierarchically structured categories to doc- uments is known as hierarchical multi-label text classification (HMTC). For the hierarchical multi-label classification problem, it is essential to model the dependencies among categories in the hierarchical structure. The vertical correlations capture the dependencies of categories at different levels, and the horizontal correlations reflect the relationships at the same level. Straightforwardly, the HMTC problem can be reduced to a flat multi-label problem (Fall et al., 2003), which simply ignores the vertical category correlations. To address that, attempts have been made to exploit the hierarchical dependencies to improve the classification performance. Among them, (Costa et al., 2007) adopts a top-down way to generate a hierarchical structure of decision-tree- based classifiers to predict categories at the cor- responding hierarchical level. (Wehrmann et al., 2018) proposes a unified framework that combines the local outputs of each category hierarchical level …

Content truncated to 500 words