Counterspeech has proven to be a powerful tool to combat hate speech online.Previous studies have focused on generating counterspeech conditioned only on specific strategies (single attributed).However, a holistic approach considering multiple attributes simultaneously can yield more nuanced and effective responses.Here, we introduce HiPPrO, Hierarchical Prefix learning with Preference Optimization, a novel two-stage framework that utilizes the effectiveness of attribute-specific prefix embedding spaces hierarchically optimized during the counterspeech generation process in the first phase.Thereafter, we incorporate both reference and reward-free preference optimization to generate more constructive counterspeech.Furthermore, we extend IntentCONANv2 by annotating all 13, 973 counterspeech instances with emotion labels by five annotators.HiPPrO leverages hierarchical prefix optimization to integrate these dual attributes effectively.An extensive evaluation demonstrates that HiPPrO achieves a ∼ 38% improvement in strategy conformity and a ∼ 3%, ∼ 2%, ∼ 3% improvement in Rouge-1, Rouge-2, and Rouge-L, respectively, compared to several baseline models.Human evaluations further substantiate the superiority of our approach, highlighting the enhanced relevance and appropriateness of the generated counterspeech.This work underscores the potential of multiattribute conditioning in advancing the efficacy of counterspeech generation systems. 1 Our code is available on Github and dataset is opensourced on Hugging-face.

Amey Hengle, Aswini Kumar Padhi, Anil Bandhakavi, Tanmoy Chakraborty. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Amey Hengle, Aswini Padhi, Sahajpreet Singh, Anil Bandhakavi, Md Shad Akhtar, Tanmoy Chakraborty. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Rishabh Gupta, Shaily Desai, Manvi Goel, Anil Bandhakavi, Tanmoy Chakraborty, Md. Shad Akhtar. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Given a tweet, predicting the discussions that unfold around it is convoluted, to say the least. Most if not all of the discernibly benign tweets which seem innocuous may very well attract inflammatory posts (hate speech) from people who find them non-congenial. Therefore, building upon the aforementioned task and predicting if a tweet will incite hate speech is of critical importance. To stifle the dissemination of online hate speech is the need of the hour. Thus, there have been a handful of models for the detection of hate speech. Classical models work retrospectively by leveraging a reactive strategy – detection after the postage of hate speech, i.e., a backward trace after detection. Therefore, a benign post that may act as a surrogate to invoke toxicity in the near future, may not be flagged by the existing hate speech detection models. In this paper, we address this problem through a proactive strategy initiated to avert hate crime. We propose DRAGNET, a deep stratified learning framework which predicts the intensity of hatred that a root tweet can fetch through its subsequent replies. We extend the collection of social media discourse from our earlier work [1], comprising the entire reply chains up to $\sim$5k root tweets catalogued into four controversial topics Similar to [1], we notice a handful of cases where despite the root tweets being non-hateful, the succeeding replies inject an enormous amount of toxicity into the discussions. DRAGNET turns out to be highly effective, significantly outperforming six state-of-the-art baselines. It beats the best baseline with an increase of 9.4% in the Pearson correlation coefficient and a decrease of 19% in Root Mean Square Error. Further, DRAGNET’S deployment in Logically’s advanced AI platform designed to monitor real-world problematic and hateful narratives has improved the aggregated insights extracted for understanding their spread, influence and thereby offering actionable intelligence to counter them

Curbing hate speech is undoubtedly a major challenge for online microblogging platforms like Twitter. While there have been studies around hate speech detection, it is not clear how hate speech finds its way into an online discussion. It is important for a content moderator to not only identify which tweet is hateful but also to predict which tweet will be responsible for accumulating hate speech. This would help in prioritizing tweets that need constant monitoring. Our analysis reveals that for hate speech to manifest in an ongoing discussion, the source tweet may not necessarily be hateful; rather, there are plenty of such non-hateful tweets which gradually invoke hateful replies, resulting in the entire reply threads becoming provocative.

Abstract Theoretical frameworks in psychology map the relationships between emotions and sentiments. In this paper, we study the role of such mapping for computational emotion detection from text (e.g., social media) with an aim to understand the usefulness of an emotion‐rich corpus of documents (e.g., tweets) to learn polarity lexicons for sentiment analysis. We propose two different methods that leverage a corpus of emotion‐labelled tweets to learn word‐polarity lexicons. The proposed methods model the emotion corpus using a generative unigram mixture model, combined with the emotion‐sentiment mapping proposed in psychology for automated generation of word‐polarity lexicons that capture emotion‐rich vocabulary. We comparatively evaluate the quality of the proposed mixture model in learning emotion‐aware sentiment lexicons with those generated using supervised latent dirichlet allocation (sLDA) and word‐document‐frequency (WDF) statistics. Sentiment analysis experiments on benchmark Twitter data sets confirm the quality of our proposed lexicons. Further, a comparative analysis with sLDA, WDF‐based emotion‐aware lexicons, and standard sentiment lexicons that are agnostic to emotion knowledge suggests that the proposed lexicons lead to a significantly better performance in both sentiment classification and sentiment intensity prediction tasks.