Hierarchical Text Classification with Reinforced Label Assignment

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 445–455, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 445 Hierarchical Text Classification with Reinforced Label Assignment Yuning Mao1, Jingjing Tian2, Jiawei Han1, Xiang Ren3 1Department of Computer Science, University of Illinois at Urbana-Champaign, IL, USA 3Department of Computer Science, Peking University, Beijing, China 3Department of Computer Science, University of Southern California, CA, USA 1{yuningm2, hanj}@illinois.edu 2tianjj97@pku.edu.cn 3xiangren@usc.edu Abstract While existing hierarchical text classification (HTC) methods attempt to capture label hier- archies for model training, they either make local decisions regarding each label or com- pletely ignore the hierarchy information dur- ing inference. To solve the mismatch between training and inference as well as modeling la- bel dependencies in a more principled way, we formulate HTC as a Markov decision pro- cess and propose to learn a Label Assignment Policy via deep reinforcement learning to de- termine where to place an object and when to stop the assignment process. The proposed method, HiLAP, explores the hierarchy dur- ing both training and inference time in a con- sistent manner and makes inter-dependent de- cisions. As a general framework, HiLAP can incorporate different neural encoders as base models for end-to-end training. Experiments on five public datasets and four base models show that HiLAP yields an average improve- ment of 33.4% in Macro-F1 over flat classifiers and outperforms state-of-the-art HTC methods by a large margin.1 1 Introduction In recent years there has been a surge of interest in leveraging hierarchies (taxonomies) to organize objects (e.g., documents), leading to the develop- ment of hierarchical text classification (HTC)—a task that aims to predict for an object multiple ap- propriate labels in a given label hierarchy, which together constitute a sub-tree. HTC methods have found a wide range of applications such as ques- tion answering (Qu et al., 2012), online advertis- ing (Agrawal et al., 2013), and scientific litera- ture organization (Peng et al., 2016). In contrast to “flat” classification, the key challenges of HTC 1Data and code can be found at https://github.com/ morningmoni/HiLAP. X ROOT Restaurants Caribbean American Italian Bars Nightlife Beer Bars Wine Bars Dominican Haitian Puerto Rican Label Hierarchy Inconsistent Prediction Single-Path Prediction Mandatory Leaf Node Prediction Non-mandatory Leaf Node Prediction Yelp Business X X ROOT Restaurants Caribbean American Italian Bars Nightlife Beer Bars Wine Bars Dominican Haitian Puerto Rican X ROOT Restaurants Caribbean American Italian Bars Nightlife Beer Bars Wine Bars Dominican Haitian Puerto Rican ROOT Restaurants Caribbean American Italian Bars Nightlife Beer Bars Wine Bars Dominican Haitian Puerto Rican Figure 1: We aim at consistent, multi-path, and non- mandatory leaf node prediction. For a Caribbean restaurant with a beer bar, inconsistent prediction may place it to node “Beer Bars” but not “Bars”, which con- tradicts with each other; Single-path prediction may only recognize that it is a beer bar; Mandatory leaf node prediction would have to assign a leaf node “Domini- can” even if the …

Content truncated to 500 words