Categorizing Offensive Language in Social Networks: A Chinese Corpus, Systems and an Explainable Tool

Categorizing Offensive Language in Social Networks: A Chinese Corpus, Systems and an Explanation Tool Xiangru Tang, Xianjun Shen∗, Yujie Wang, Yujuan Yang School of Computer, Central China Normal University, China National Language Resources Monitoring & Research Center for Network Media, China Hubei Provincial Key Laboratory of Artificial Intelligence and Smart Learning, China xjshen@mail.ccnu.edu Abstract Recently, more and more data have been generated in the online world, filled with offensive lan- guage such as threats, swear words or straightforward insults. It is disgraceful for a progressive society, and then the question arises on how language resources and technologies can cope with this challenge. However, previous work only analyzes the problem as a whole but fails to de- tect particular types of offensive content in a more fine-grained way, mainly because of the lack of annotated data. In this work, we present a densely annotated data-set COLA (Categorizing Offensive LAnguage), consists of fine-grained insulting language, antisocial language and ille- gal language. We study different strategies for automatically identifying offensive language on COLA data. Further, we design a capsule system with hierarchical attention to aggregate and fully utilize information, which obtains a state-of-the-art result. Results from experiments prove that our hierarchical attention capsule network (HACN) performs significantly better than exist- ing methods in offensive classification with the precision of 94.37% and recall of 95.28%. We also explain what our model has learned with an explanation tool called Integrated Gradients. Meanwhile, our system’s processing speed can handle each sentence in 10msec, suggesting the potential for efficient deployment in real situations. 1 Introduction In modern society, the occupation of offensive language on the online world, such as social media, is be- coming a paramount concern. Offensive language differs considerably, ranging from pure abuse to more rigorous types of writing. Thus, offensive language is hard to be automatically identified. However, it’s essential to track this; for example, the appearance of offensive language on social media is related to hate crimes in a real social situation. [M¨uller and Schwarz2018]. Moreover, it can be pretty trouble- some to distinguish fine-grained offensive language because few general definitions exist [Davidson et al.2017]. Recently, researchers have proposed some guidelines to identify the type and the attributes of offensive language [Zampieri et al.2019a]. However, the online world’s offensive language is a general category containing specific examples of profanity or insult. In our work, “Offensive language” in the online world is defined in more detail and fine-grained. And to the best of our knowledge, though offensive language identification being a burgeoning field, there is no data-set yet for Chinese. “Offensive” is pretty much something people identify as against morals, very inappropriate, or disre- spectful. However, “offensive” is a broad general term and does not define the precise extent or the limits of its application. Thus, we classify the term “offensive” into three categories: “insulting,” “antisocial” and “illegal” through stepwise refinement. “Insulting” is something rude, insensitive and/or offensive, directed at another person or group of people. This emphasizes that the content is a direct …

Content truncated to 500 words