Peer-Label Assisted Hierarchical Text Classification

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3747–3758 July 9-14, 2023 ©2023 Association for Computational Linguistics Peer-Label Assisted Hierarchical Text Classification Junru Song1, Feifei Wang2,3∗, Yang Yang4 1Institute of Statistics and Big Data, Renmin University of China 2Center for Applied Statistics, Renmin University of China 3School of Statistics, Renmin University of China 4Defense Innovation Institute, Chinese Academy of Military Science {songjunru,feifei.wang}@ruc.edu.cn, bigyangy@gmail.com Abstract Hierarchical text classification (HTC) is a chal- lenging task, in which the labels of texts can be organized into a category hierarchy. To deal with the HTC problem, many existing works focus on utilizing the parent-child relationships that are explicitly shown in the hierarchy. How- ever, texts with a category hierarchy also have some latent relevancy among labels in the same level of the hierarchy. We refer to these labels as peer labels, from which the peer effects are originally utilized in our work to improve the classification performance. To fully explore the peer-label relationship, we develop a PeerHTC method. This method innovatively measures the latent relevancy of peer labels through sev- eral metrics and then encodes the relevancy with a Graph Convolutional Neural Network. We also propose a sample importance learning method to ameliorate the side effects raised by modelling the peer label relevancy. Our experi- ments on several standard datasets demonstrate the evidence of peer labels and the superiority of PeerHTC over other state-of-the-art HTC methods in terms of classification accuracy. 1 Introduction Hierarchical text classification (HTC) is a multi- label text classification problem which aims to clas- sify texts into categories that can be organized into a taxonomic hierarchy. It is an important problem in natural language processing and has attracted in- creasing attention in both industrial and academic fields. Typical HTC problems include patent cat- egorization (Gomez and Moens, 2014), medical record coding (Cao et al., 2020), and product cate- gorization (Cevahir and Murakami, 2016). Due to the complexity of category hierarchy, the problem of hierarchical text classification is more challenging than plain text classification. The parent-child relationships between categories in ad- jacent levels of the hierarchy are usually defined ∗ Corresponding author. in advance. Then a natural way to solve the HTC problem is to incorporate this prior knowledge into the model, i.e., making the model aware of the hierarchy. Building "hierarchy-aware" models is beneficial for HTC, which is particularly true for categories with few samples. Therefore, it has long been the main focus in HTC to figure out the most effective way of utilizing the category hierarchy to improve the classification performance. In the past literature, existing approaches for HTC can be generally categorized into three groups: local approaches, global approaches, and local-global-combined ones (also known as hybrid approaches). The local approaches train local clas- sifiers for every child label, every parent label or every level in the hierarchy (Shimura et al., 2018; Banerjee et al., 2019). The parameters of local classifiers are initialized in a top-down fashion ac- cording to …

Content truncated to 500 words