Generation-Based Data Augmentation for Offensive Language Detection: Is It Worth It?

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3359–3377 May 2-6, 2023 ©2023 Association for Computational Linguistics Generation-Based Data Augmentation for Offensive Language Detection: Is It Worth It? Camilla Casula♢♣, Sara Tonelli♢ ♢Fondazione Bruno Kessler, Trento, Italy ♣University of Trento, Italy {ccasula,satonelli}@fbk.eu Abstract Generation-based data augmentation (DA) has been presented in several works as a way to improve offensive language detection. How- ever, the effectiveness of generative DA has been shown only in limited scenarios, and the potential injection of biases when using gener- ated data to classify offensive language has not been investigated. Our aim is that of analyzing the feasibility of generative data augmentation more in-depth with two main focuses. First, we investigate the robustness of models trained on generated data in a variety of data augmenta- tion setups, both novel and already presented in previous work, and compare their perfor- mance on four widely-used English offensive language datasets that present inherent differ- ences in terms of content and complexity. In addition to this, we analyze models using the HateCheck suite, a series of functional tests created to challenge hate speech detection sys- tems. Second, we investigate potential lexical bias issues through a qualitative analysis of the generated data. We find that the potential posi- tive impact of generative data augmentation on model performance is unreliable, and genera- tive DA can also have unpredictable effects on lexical bias. � Warning: this paper contains exam- ples that may be offensive or upsetting. 1 Introduction Even though large language models have been found to have a tendency to encode and propa- gate undesirable social bias (Bender et al., 2021), going as far as generating toxic sequences start- ing from non-toxic prompts (Gehman et al., 2020), the use of synthetic data for offensive language detection has been found to be potentially help- ful in improving models (e.g. Juuti et al. (2020); Wullach et al. (2021); D’Sa et al. (2021)). Indeed, data augmentation (DA) through generation has the potential to mitigate some of the known issues of smaller datasets, which are common in offensive language detection, such as lack of linguistic varia- tion and risk of overfitting (Vidgen and Derczynski, 2020). Furthermore, synthetic data can overcome privacy issues related to the use of social media data obtained without user consent in research ex- periments. It can also mitigate dataset decay, an issue affecting reproducibility, since online mes- sages, especially abusive ones, tend to be deleted over time, while synthetic examples do not present this issue (Klubicka and Fernández, 2018). While generative DA has been shown to be po- tentially useful for the task of detecting offensive and abusive language online in multiple works, sev- eral aspects and implications of it remain unex- plored. First of all, generative DA has mostly been shown to work for offensive language detection when starting with a single specific dataset and using a specific generation setup, with no investi- gation of the impact of different generation setups on the …

Content truncated to 500 words