Data Augmentation with Dual Training for Offensive Span Detection

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2569 - 2575 July 10-15, 2022 ©2022 Association for Computational Linguistics Data Augmentation with Dual Training for Offensive Span Detection Nasim Nouri Raouf Medical Group Tehran, Iran nasimnouri@raoufmed.com Abstract Recognizing offensive text is an important re- quirement for every content management sys- tem, especially for social networks. While the majority of the prior work formulate this problem as text classification, i.e., if a text ex- cerpt is offensive or not, in this work we pro- pose a novel model for offensive span detection (OSD), whose goal is to identify the spans re- sponsible for the offensive tone of the text. One of the challenges to train a model for this novel setting is the lack of enough training data. To address this limitation, in this work we propose a novel method in which the large-scale pre- trained language model GPT-2 is employed to generate synthetic training data for OSD. In par- ticular, we propose to train the GPT-2 model in a dual-training setting using the REINFORCE algorithm to generate in-domain, natural and di- verse training samples. Extensive experiments on the benchmark dataset for OSD reveal the effectiveness of the proposed method. 1 Introduction It’s no secret that social networks are growing in popularity. However, growth in popularity also brings some challenges, including the toxicity asso- ciated with the content posted by users. It may take different forms in social media, including insults, mockery, threats, discrimination, or swearing. The presence of offensive text in social networks can have a detrimental effect on their users, making it desirable to identify and remove them from the text. Since this is an important requirement, the task of offensive language detection has been exten- sively studied in NLP community (Schmidt and Wiegand, 2017; Wulczyn et al., 2017; Feng et al., 2018; Borkan et al., 2019; Pavlopoulos et al., 2019; Sivanaiah et al., 2020; Yasaswini et al., 2021) Most existing works, however, only classify a text snip- pet as offensive or not, failing to provide further information on which specific words and phrases in the text contribute the most to its offensive tone. If the text snippet is lengthy, the moderators will need this information to decide how to proceed with the offenses flagged. As such, in this work, we fill this gap by proposing a novel model for the task of offensive span detection (OSD). As an example, in the given text “This live streamer clearly has no brain; he is such a tool!", the phrase “has no brain" and the slang word “tool" are two offensive spans responsible for the toxicity of the text. One of the barriers to this task is the lack of labeled data. Inspired by the recent advances in the appli- cation of pre-trained language models to augment training data for low-resources tasks (Zhang et al., 2020; Yang et al., 2020; Peng et al., 2020; Kumar et al., 2020; Anaby-Tavor …

Content truncated to 500 words