NLP_UIOWA at Semeval-2021 Task 5: Transferring Toxic Sets to Tag Toxic Spans

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 881–887 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 881 NLP UIOWA at Semeval-2021 Task 5: Transferring Toxic Sets to Tag Toxic Spans Jonathan Rusert Department of Computer Science University of Iowa Iowa City, IA, USA jonathan-rusert@uiowa.edu Abstract We leverage a BLSTM with attention to iden- tify toxic spans in texts. We explore differ- ent dimensions which affect the model’s per- formance. The first dimension explored is the dataset the model is trained on. Besides the provided dataset, we explore the transferabil- ity of 5 different toxic related sets, including offensive, toxic, abusive, and hate sets. We find that the solely offensive set shows the highest promise of transferability. The sec- ond dimension we explore is methodology, including leveraging attention, employing a greedy remove method, using a frequency ra- tio, and examining hybrid combinations of multiple methods. We conduct an error analy- sis to examine which types of toxic spans were missed and which were wrongly inferred as toxic along with the main reasons why they oc- curred. Finally, we extend our method via en- sembles, which achieves our highest F1 score of 55.1. 1 Introduction Social media platforms (e.g. Twitter) benefit from a welcoming environment. Toxicity introduced by users can drive new (or existing) users away. There- fore, it is beneficial for platforms to moderate the amount of toxic language allowed. Often human moderators are employed to address this. How- ever, as the numbers of users increase on a plat- form, it can be difficult for human moderation as the sheer amount of content generated increases as well. Thus platforms employ automated assistance (Terdiman, 2018). While automated toxic language classifiers have made strides at determining if text is toxic (Yenala et al., 2018; Zhang et al., 2018), much work is still left for humans who are reviewing the text. In large texts, the reason for toxicity might only be one or two words, however, the moderator will have to Text Toxic Span What a jerk! [7, 8, 9, 10] “Trump is up to the task” LMAO, you are an idiot [44, 45, 46, 47, 48] Canada: Institution- alised mediocrity. [ ] Contrary to popular be- lief not all criminals are stupid. Fortunately, there are plenty like this moron that are. [49, 50, 51, 52, 53, 54, 98, 99, 100, 101, 102] Table 1: Examples of toxic spans in toxic texts. The numbers indicate the character position of the toxic span. Empty brackets indicate no specific span. scan the entire text to find the toxicity. Repeating this process across many texts is cumbersome. Au- tomated systems could assist human moderators further by noting which areas in the text are toxic, thus allowing the human to quickly check text for toxicity. Finding these toxic areas is the task that is proposed by (Pavlopoulos et al., 2021). Specif- ically, given English toxic text, determine where the “toxic spans” occur, where a “toxic span” is the character positions in …

Content truncated to 500 words