UTFPR at SemEval-2019 Task 6: Relying on Compositionality to Find Offense

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 801–805 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 801 UTFPR at SemEval-2019 Task 6: Relying on Compositionality to Find Offense Gustavo Henrique Paetzold Universidade Tecnol´ogica Federal do Paran´a / Toledo-PR, Brazil ghpaetzold@utfpr.edu.br Abstract We present the UTFPR system for the Of- fensEval shared task of SemEval 2019: A character-to-word-to-sentence compositional RNN model trained exclusively over the train- ing data provided by the organizers. We find that, although not very competitive for the task at hand, it offers a robust solution to the ortho- graphic irregularity inherent to tweets. 1 Introduction Text classification tasks can take a wide variety of forms, and some of them, such as sentiment and emotion analysis, have managed to grab a lot of attention from researchers in recent years. More recently, however, the public’s growing engage- ment in debates on the topics of free speech and politics has led the Natural Language Processing (NLP) and Machine Learning (ML) communities to take an interest in classification tasks related to identifying and categorizing patterns of profan- ity, hate speech and offense. The prominence of shared tasks held on these topics, such as those of Fersini et al. (2018) and Wiegand et al. (2018), are great examples of the research community com- ing together to attempt to create more reliable so- lutions for these challenges. While hate speech is commonly characterized by specific slurs and other offensive expressions that convey prejudice against a certain group or individual, offensive speech is often more chal- lenging to identify because it encompasses a more broad spectrum of language, featuring expressions that do not necessarily convey prejudice (Malmasi and Zampieri, 2018). Technologies that identify these types of patterns could, for instance, help a social media platform on profiling users and take appropriate action whenever someone breaks user agreements and/or terms of use. Identifying offensive language within the con- tent of social media platforms is particularly chal- lenging, since this type of content is usually lit- tered with irregular orthography, meta-characters, slang and others. Since a lot of the effort from the research community focuses on identifying offen- sive language in social media platforms, an NLP approach for such a task must be able to overcome those hurdles in some way. Some of the preferred methods for handling the orthographic irregular- ity of social media content are using word embed- dings trained over tweets (Rozental et al., 2018) or regularizing unusual spellings (Bertaglia and das Grac¸as Volpe Nunes, 2017), but neither of them ensure that every possible orthographically irreg- ular word will be understood by the NLP model in question. Recently, however, there have been a lot of contributions that present compositional neural models that learn numerical representations of words based on the sequence of characters that compose them (Kim et al., 2016; Ling et al., 2015; Balazs et al., 2018; Paetzold, 2018). These mod- els have been demonstrated to be both effective in text classification, and robust when …

Content truncated to 500 words