Comparison of Machine Learning Approaches for Industry Classification Based on Textual Descriptions of Companies

Proceedings of Recent Advances in Natural Language Processing, pages 1169–1175, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_134 1169 Comparison of Machine Learning Approaches for Industry Classification Based on Textual Descriptions of Companies Andrey Tagarev1 and Nikola Tulechki1 and Svetla Boytcheva1,2 1Sirma AI trading as Ontotext, Bulgara 2Institute of Information and Communication Technologies, Bulgarian Academy of Sciences {andrey.tagarev,nikola.tulechki,svetla.boytcheva}@ontotext.com Abstract This paper addresses the task of catego- rizing companies within industry classi- fication schemes. The dataset consists of encyclopedic articles about companies and their economic activities. The tar- get classification schema is build by map- ping linked open data in a semi-supervised manner. Target classes are built bottom-up from DBpedia. We apply several state of the art text classification techniques, based both on deep learning and classical vector- space models. 1 Motivation The era of big data has made the task of integrat- ing several heterogeneous sources quite common and important. This is a challenging task because, typically, data representation of the same object can significantly differ in different sources, both in level of detail and type of available informa- tion. In addition, different ontologies (classifica- tion schema) are used for the same concepts in dif- ferent datasets. Ontology mapping is itself a chal- lenge even for human experts. Thus the problem of concept classification is very important in data integration. This paper presents a comparison of different techniques for solving the task of company indus- try classification based on textual descriptions of companies in DBpedia1. The problem of text- classification is defined as follows: for a collection of company descriptions D = {d1, d2, ..., dn}, assign one or more industry categories to each company from a discrete set of labels C = {c1, c2, ..., ck}. There are more than fifteen different com- pany industry classifications of varying granu- 1https://wiki.dbpedia.org larity. Some of the most popular are: Indus- try Classification Benchmark (ICB)2, Global In- dustry Classification Standard (GICS)3, Thomson Reuters Business Classification (TRBC)4, and In- ternational Standard Industrial Classification of All Economic Activities (ISIC)5. 2 Text Classification Methods Text classification methods are widely used in sev- eral applications like e-mail spam filtering (Youn and McLeod, 2007), news categorization (Lin and Hauptmann, 2002), in marketing for product re- views (Dang et al., 2009), etc. The classical methods (Aggarwal and Zhai, 2012) in text classification are based on standard techniques. Usually the texts from the training corpus are transformed into vectors, where distinct words represent features. One of the simplest and most efficient methods is the probabilistic classi- fier Na¨ıve Bayes (NB) that has really good per- formance even with few examples and sparse fea- tures and works with a ”na¨ıve” assumption for at- tributes that are conditionally independent. How- ever, usually not all words in a text are indepen- dent. Many techniques have been developed (Al- Aidaroos et al., 2010) to overcome the problems caused by the existence of attribute correlations that can lead to classification bias and negatively affect an NB classifier’s performance. McCallum and Nigam (McCallum et al., 1998) demonstrate …

Content truncated to 500 words