Counterspeeches up my sleeve! Intent Distribution Learning and Persistent Fusion for Intent-Conditioned Counterspeech Generation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5792–5809 July 9-14, 2023 ©2023 Association for Computational Linguistics Counterspeeches up my sleeve! Intent Distribution Learning and Persistent Fusion for Intent-Conditioned Counterspeech Generation Rishabh Gupta1, Shaily Desai1, Manvi Goel1, Anil Bandhkavi2, Tanmoy Chakraborty3, and Md Shad Akhtar1 1IIIT Delhi, India, 2Logically, U.K., 3IIT Delhi, India {rishabh19089, shailyd, manvi19472, shad.akhtar}@iiitd.ac.in, anil@logically.ai, tanchak@iitd.ac.in Abstract Counterspeech has been demonstrated to be an efficacious approach for combating hate speech. While various conventional and con- trolled approaches have been studied in recent years to generate counterspeech, a counter- speech with a certain intent may not be suf- ficient in every scenario. Due to the complex and multifaceted nature of hate speech, utiliz- ing multiple forms of counter-narratives with varying intents may be advantageous in dif- ferent circumstances. In this paper, we ex- plore intent-conditioned counterspeech genera- tion. At first, we develop IntentCONAN, a di- versified intent-specific counterspeech dataset with 6831 counterspeeches conditioned on five intents, i.e., informative, denouncing, ques- tion, positive, and humour. Subsequently, we propose QUARC, a two-stage framework for intent-conditioned counterspeech genera- tion. QUARC leverages vector-quantized rep- resentations learned for each intent category along with PerFuMe, a novel fusion module to incorporate intent-specific information into the model. Our evaluation demonstrates that QUARC outperforms several baselines by an av- erage of ~10% across evaluation metrics. An extensive human evaluation supplements our hypothesis of better and more appropriate re- sponses than comparative systems. Warning: This work contains offensive and hateful text that some might find upsetting. It does not repre- sent the views of the authors. 1 Introduction The quantity and accessibility of information on the Internet are constantly growing in the 21st century. This has made it increasingly simpler for users on social media to post hateful or attacking speech, all while hiding behind the veil of anonymity (Mon- dal et al., 2017). Hate speech (Awal et al., 2021; Chakraborty and Masud, 2022) is an offensive dia- logue that uses stereotypes to communicate a hate- ful ideology, and it can target several protected HS: Do you have any proof that the Holocaust happened? Intent-conditioned CS generation – QUARC CS: The physical evidence of the concentration camps are undeniable. CS: How can you deny the deaths of more than 6 million Jews? CS: Historians agree that the concentration camps were a great example of a systematic persecution of Jews by the Nazi regime occurred. CS: It is a myth perpetuated by anti-Semites to justify their hatred of Jews. (Detected Intent: Denouncing) CS: Do you have any proof on your claim about hate? (Detected Intent: Question) GPS DialoGPT Adequate, but the informative counter argument would be more effective. Classical CS generation Denouncing Question Informative Figure 1: Outputs compared to pre-existing methods. These examples show different intents generated by different models. This raises the need for a system that, along with producing multiple counter-arguments, also ensures that the generated sentence is effective. qualities such as gender, religion, colour, and dis- …

Content truncated to 500 words