Hierarchical Multi-Label Classification of Scientific Documents

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8923 - 8937 December 7-11, 2022 ©2022 Association for Computational Linguistics Hierarchical Multi-Label Classification of Scientific Documents Mobashir Sadat Cornelia Caragea Computer Science University of Illinois Chicago msadat3@uic.edu cornelia@uic.edu Abstract Automatic topic classification has been stud- ied extensively to assist managing and index- ing scientific documents in a digital collection. With the large number of topics being avail- able in recent years, it has become necessary to arrange them in a hierarchy. Therefore, the au- tomatic classification systems need to be able to classify the documents hierarchically. In ad- dition, each paper is often assigned to more than one relevant topic. For example, a paper can be assigned to several topics in a hierarchy tree. In this paper, we introduce a new dataset for hierarchical multi-label text classification (HMLTC) of scientific papers called SciHTC, which contains 186, 160 papers and 1, 233 cat- egories from the ACM CCS tree. We estab- lish strong baselines for HMLTC and propose a multi-task learning approach for topic clas- sification with keyword labeling as an auxil- iary task. Our best model achieves a Macro- F1 score of 34.57% which shows that this dataset provides significant research opportu- nities on hierarchical scientific topic classifica- tion. We make our dataset and code available on Github.1 1 Introduction With the exponential increase of scientific docu- ments being published every year, the difficulty in managing and categorizing them in a digital col- lection is also increasing. While the enormity of the number of papers is the most important reason behind it, the problem can also be assigned to the large number of topics. It is a very difficult task to index a paper in a digital collection when there are thousands of topics to choose from. Fortunately, the large number of topics can be arranged in a hierarchy because, except for a few general top- ics, all topics can be seen as a sub-area of another topic. After arranging the topics in a hierarchy tree, the task of categorizing a paper becomes much 1https://github.com/msadat3/SciHTC simpler since now there are only a handful of top- ics to choose from at each level of the hierarchy. However, manual assignment of topics to a large number of papers is still very difficult and expen- sive, making an automatic system of hierarchical classification of scientific documents a necessity. After arranging the topics in a hierarchy, the classification task no longer remains a multi-class classification because in multi-class classification, each paper is classified into exactly one of sev- eral mutually exclusive classes or topics. However, if the topics are arranged in a hierarchy, they no longer remain mutually exclusive: a paper assigned to a topic node a in the hierarchy also gets assigned to topic node b, where b is a node in the parental history of a. For example, if a paper is classified to the area of natural language processing (NLP), it is also assigned to the topic …

Content truncated to 500 words