Punyajoy Saha

Punyajoy Saha, Kiran Garimella, Narla Komal Kalyan et al.

Recently, social media platforms are heavily moderated to prevent the spread of online hate speech, which is usually fertile in toxic words and is directed toward an individual or a community. Owing to such heavy moderation, newer and more subtle techniques are being deployed. One of the most striking among these is fear speech. Fear speech, as the name suggests, attempts to incite fear about a target community. Although subtle, it might be highly effective, often pushing communities toward a physical conflict. Therefore, understanding their prevalence in social media is of paramount importance. This article presents a large-scale study to understand the prevalence of 400K fear speech and over 700K hate speech posts collected from Gab.com. Remarkably, users posting a large number of fear speech accrue more followers and occupy more central positions in social networks than users posting a large number of hate speech. They can also reach out to benign users more effectively than hate speech users through replies, reposts, and mentions. This connects to the fact that, unlike hate speech, fear speech has almost zero toxic content, making it look plausible. Moreover, while fear speech topics mostly portray a community as a perpetrator using a (fake) chain of argumentation, hate speech topics hurl direct multitarget insults, thus pointing to why general users could be more gullible to fear speech. Our findings transcend even to other platforms (Twitter and Facebook) and thus necessitate using sophisticated moderation policies and mass awareness to combat fear speech.

Sarthak Roy, Ashish Harshavardhan, Animesh Mukherjee et al.

Recently efforts have been made by social media platforms as well as researchers to detect hateful or toxic language using large language models. However, none of these works aim to use explanation, additional context and victim community information in the detection process. We utilise different prompt variation, input information and evaluate large language models in zero shot setting (without adding any in-context examples). We select three large language models (GPT-3.5, text-davinci and Flan-T5) and three datasets - HateXplain, implicit hate and ToxicSpans. We find that on average including the target information in the pipeline improves the model performance substantially (~20-30%) over the baseline across the datasets. There is also a considerable effect of adding the rationales/explanations into the pipeline (~10-20%) over the baseline across the datasets. In addition, we further provide a typology of the error cases where these large language models fail to (i) classify and (ii) explain the reason for the decisions they take. Such vulnerable points automatically constitute 'jailbreak' prompts for these models and industry scale safeguard techniques need to be developed to make the models robust against such prompts.

Mithun Das, Somnath Banerjee, Punyajoy Saha et al.

Social media often serves as a breeding ground for various hateful and offensive content. Identifying such content on social media is crucial due to its impact on the race, gender, or religion in an unprejudiced society. However, while there is extensive research in hate speech detection in English, there is a gap in hateful content detection in low-resource languages like Bengali. Besides, a current trend on social media is the use of Romanized Bengali for regular interactions. To overcome the existing research's limitations, in this study, we develop an annotated dataset of 10K Bengali posts consisting of 5K actual and 5K Romanized Bengali tweets. We implement several baseline models for the classification of such hateful posts. We further explore the interlingual transfer mechanism to boost classification performance. Finally, we perform an in-depth error analysis by looking into the misclassified posts by the models. While training actual and Romanized datasets separately, we observe that XLM-Roberta performs the best. Further, we witness that on joint training and few-shot training, MuRIL outperforms other models by interpreting the semantic expressions better. We make our code and dataset public for others.

Punyajoy Saha, Kanishk Singh, Adarsh Kumar et al.

Recently, many studies have tried to create generation models to assist counter speakers by providing counterspeech suggestions for combating the explosive proliferation of online hate. However, since these suggestions are from a vanilla generation model, they might not include the appropriate properties required to counter a particular hate speech instance. In this paper, we propose CounterGeDi - an ensemble of generative discriminators (GeDi) to guide the generation of a DialoGPT model toward more polite, detoxified, and emotionally laden counterspeech. We generate counterspeech using three datasets and observe significant improvement across different attribute scores. The politeness and detoxification scores increased by around 15% and 6% respectively, while the emotion in the counterspeech increased by at least 10% across all the datasets. We also experiment with triple-attribute control and observe significant improvement over single attribute results when combining complementing attributes, e.g., politeness, joyfulness and detoxification. In all these experiments, the relevancy of the generated text does not deteriorate due to the application of these controls

Mithun Das, Punyajoy Saha, Binny Mathew et al.

Due to the sheer volume of online hate, the AI and NLP communities have started building models to detect such hateful content. Recently, multilingual hate is a major emerging challenge for automated detection where code-mixing or more than one language have been used for conversation in social media. Typically, hate speech detection models are evaluated by measuring their performance on the held-out test data using metrics such as accuracy and F1-score. While these metrics are useful, it becomes difficult to identify using them where the model is failing, and how to resolve it. To enable more targeted diagnostic insights of such multilingual hate speech models, we introduce a set of functionalities for the purpose of evaluation. We have been inspired to design this kind of functionalities based on real-world conversation on social media. Considering Hindi as a base language, we craft test cases for each functionality. We name our evaluation dataset HateCheckHIn. To illustrate the utility of these functionalities , we test state-of-the-art transformer based m-BERT model and the Perspective API.

Piush Aggarwal, Pranit Chawla, Mithun Das et al.

Exploiting social media to spread hate has tremendously increased over the years. Lately, multi-modal hateful content such as memes has drawn relatively more traction than uni-modal content. Moreover, the availability of implicit content payloads makes them fairly challenging to be detected by existing hateful meme detection systems. In this paper, we present a use case study to analyze such systems' vulnerabilities against external adversarial attacks. We find that even very simple perturbations in uni-modal and multi-modal settings performed by humans with little knowledge about the model can make the existing detection models highly vulnerable. Empirically, we find a noticeable performance drop of as high as 10% in the macro-F1 score for certain attacks. As a remedy, we attempt to boost the model's robustness using contrastive learning as well as an adversarial training-based method - VILLA. Using an ensemble of the above two approaches, in two of our high resolution datasets, we are able to (re)gain back the performance to a large extent for certain attacks. We believe that ours is a first step toward addressing this crucial problem in an adversarial setting and would inspire more such investigations in the future.

Punyajoy Saha, Divyanshu Sheth, Kushal Kedia et al.

Abusive language is a concerning problem in online social media. Past research on detecting abusive language covers different platforms, languages, demographies, etc. However, models trained using these datasets do not perform well in cross-domain evaluation settings. To overcome this, a common strategy is to use a few samples from the target domain to train models to get better performance in that domain (cross-domain few-shot training). However, this might cause the models to overfit the artefacts of those samples. A compelling solution could be to guide the models toward rationales, i.e., spans of text that justify the text's label. This method has been found to improve model performance in the in-domain setting across various NLP tasks. In this paper, we propose RGFS (Rationale-Guided Few-Shot Classification) for abusive language detection. We first build a multitask learning setup to jointly learn rationales, targets, and labels, and find a significant improvement of 6% macro F1 on the rationale detection task over training solely rationale classifiers. We introduce two rationale-integrated BERT-based architectures (the RGFS models) and evaluate our systems over five different abusive language datasets, finding that in the few-shot classification setting, RGFS-based models outperform baseline models by about 7% in macro F1 scores and perform competitively to models finetuned on other source domains. Furthermore, RGFS-based models outperform LIME/SHAP-based approaches in terms of plausibility and are close in performance in terms of faithfulness.

Millon Madhur Das, Punyajoy Saha, Mithun Das

The proliferation of online hate speech has necessitated the creation of algorithms which can detect toxicity. Most of the past research focuses on this detection as a classification task, but assigning an absolute toxicity label is often tricky. Hence, few of the past works transform the same task into a regression. This paper shows the comparative evaluation of different transformers and traditional machine learning models on a recently released toxicity severity measurement dataset by Jigsaw. We further demonstrate the issues with the model predictions using explainability analysis.

Binny Mathew, Punyajoy Saha, Seid Muhie Yimam et al.

Hate speech is a challenging issue plaguing the online social media. While better models for hate speech detection are continuously being developed, there is little research on the bias and interpretability aspects of hate speech. In this paper, we introduce HateXplain, the first benchmark hate speech dataset covering multiple aspects of the issue. Each post in our dataset is annotated from three different perspectives: the basic, commonly used 3-class classification (i.e., hate, offensive or normal), the target community (i.e., the community that has been the victim of hate speech/offensive speech in the post), and the rationales, i.e., the portions of the post on which their labelling decision (as hate, offensive or normal) is based. We utilize existing state-of-the-art models and observe that even models that perform very well in classification do not score high on explainability metrics like model plausibility and faithfulness. We also observe that models, which utilize the human rationales for training, perform better in reducing unintended bias towards target communities. We have made our code and dataset public at https://github.com/punyajoy/HateXplain

Sai Saketh Aluru, Binny Mathew, Punyajoy Saha et al.

Hate speech detection is a challenging problem with most of the datasets available in only one language: English. In this paper, we conduct a large scale analysis of multilingual hate speech in 9 languages from 16 different sources. We observe that in low resource setting, simple models such as LASER embedding with logistic regression performs the best, while in high resource setting BERT based models perform better. In case of zero-shot classification, languages such as Italian and Portuguese achieve good results. Our proposed framework could be used as an efficient solution for low-resource languages. These models could also act as good baselines for future multilingual hate speech detection tasks. We have made our code and experimental settings public for other researchers at https://github.com/punyajoy/DE-LIMIT.

Punyajoy Saha, Binny Mathew, Pawan Goyal et al.

Reducing hateful and offensive content in online social media pose a dual problem for the moderators. On the one hand, rigid censorship on social media cannot be imposed. On the other, the free flow of such content cannot be allowed. Hence, we require efficient abusive language detection system to detect such harmful content in social media. In this paper, we present our machine learning model, HateMonitor, developed for Hate Speech and Offensive Content Identification in Indo-European Languages (HASOC), a shared task at FIRE 2019. We have used a Gradient Boosting model, along with BERT and LASER embeddings, to make the system language agnostic. Our model came at First position for the German sub-task A. We have also made our model public at https://github.com/punyajoy/HateMonitors-HASOC .

Punyajoy Saha, Binny Mathew, Pawan Goyal et al.

With the online proliferation of hate speech, there is an urgent need for systems that can detect such harmful content. In this paper, We present the machine learning models developed for the Automatic Misogyny Identification (AMI) shared task at EVALITA 2018. We generate three types of features: Sentence Embeddings, TF-IDF Vectors, and BOW Vectors to represent each tweet. These features are then concatenated and fed into the machine learning models. Our model came First for the English Subtask A and Fifth for the English Subtask B. We release our winning model for public use and it's available at https://github.com/punyajoy/Hateminers-EVALITA.

Punyajoy Saha, Aalok Agrawal, Abhik Jana et al.

With the emergence of numerous Large Language Models (LLM), the usage of such models in various Natural Language Processing (NLP) applications is increasing extensively. Counterspeech generation is one such key task where efforts are made to develop generative models by fine-tuning LLMs with hatespeech - counterspeech pairs, but none of these attempts explores the intrinsic properties of large language models in zero-shot settings. In this work, we present a comprehensive analysis of the performances of four LLMs namely GPT-2, DialoGPT, ChatGPT and FlanT5 in zero-shot settings for counterspeech generation, which is the first of its kind. For GPT-2 and DialoGPT, we further investigate the deviation in performance with respect to the sizes (small, medium, large) of the models. On the other hand, we propose three different prompting strategies for generating different types of counterspeech and analyse the impact of such strategies on the performance of the models. Our analysis shows that there is an improvement in generation quality for two datasets (17%), however the toxicity increase (25%) with increase in model size. Considering type of model, GPT-2 and FlanT5 models are significantly better in terms of counterspeech quality but also have high toxicity as compared to DialoGPT. ChatGPT are much better at generating counter speech than other models across all metrics. In terms of prompting, we find that our proposed strategies help in improving counter speech generation across all the models.

Naquee Rizwan, Paramananda Bhaskar, Mithun Das et al.

There is a rapid increase in the use of multimedia content in current social media platforms. One of the highly popular forms of such multimedia content are memes. While memes have been primarily invented to promote funny and buoyant discussions, malevolent users exploit memes to target individuals or vulnerable communities, making it imperative to identify and address such instances of hateful memes. Thus social media platforms are in dire need for active moderation of such harmful content. While manual moderation is extremely difficult due to the scale of such content, automatic moderation is challenged by the need of good quality annotated data to train hate meme detection algorithms. This makes a perfect pretext for exploring the power of modern day vision language models (VLMs) that have exhibited outstanding performance across various tasks. In this paper we study the effectiveness of VLMs in handling intricate tasks such as hate meme detection in a completely zero-shot setting so that there is no dependency on annotated data for the task. We perform thorough prompt engineering and query state-of-the-art VLMs using various prompt types to detect hateful/harmful memes. We further interpret the misclassification cases using a novel superpixel based occlusion method. Finally we show that these misclassifications can be neatly arranged into a typology of error classes the knowledge of which should enable the design of better safety guardrails in future.

Mithun Das, Saurabh Kumar Pandey, Shivansh Sethi et al.

With the rise of online abuse, the NLP community has begun investigating the use of neural architectures to generate counterspeech that can "counter" the vicious tone of such abusive speech and dilute/ameliorate their rippling effect over the social network. However, most of the efforts so far have been primarily focused on English. To bridge the gap for low-resource languages such as Bengali and Hindi, we create a benchmark dataset of 5,062 abusive speech/counterspeech pairs, of which 2,460 pairs are in Bengali and 2,602 pairs are in Hindi. We implement several baseline models considering various interlingual transfer mechanisms with different configurations to generate suitable counterspeech to set up an effective benchmark. We observe that the monolingual setup yields the best performance. Further, using synthetic transfer, language models can generate counterspeech to some extent; specifically, we notice that transferability is better when languages belong to the same language family.

Naquee Rizwan, Seid Muhie Yimam, Daryna Dementieva et al.

Despite regulations imposed by nations and social media platforms, e.g. (Government of India, 2021; European Parliament and Council of the European Union, 2022), inter alia, hateful content persists as a significant challenge. Existing approaches primarily rely on reactive measures such as blocking or suspending offensive messages, with emerging strategies focusing on proactive measurements like detoxification and counterspeech. In our work, which we call HatePRISM, we conduct a comprehensive examination of hate speech regulations and strategies from three perspectives: country regulations, social platform policies, and NLP research datasets. Our findings reveal significant inconsistencies in hate speech definitions and moderation practices across jurisdictions and platforms, alongside a lack of alignment with research efforts. Based on these insights, we suggest ideas and research direction for further exploration of a unified framework for automated hate speech moderation incorporating diverse strategies.

Punyajoy Saha, Sudipta Halder, Debjyoti Mondal et al.

Safety alignment is critical for deploying large language models (LLMs) in real-world applications, yet most existing approaches rely on large human-annotated datasets and static red-teaming benchmarks that are costly, difficult to scale, and slow to adapt to evolving model behaviors. Moreover, overly conservative safety mechanisms can reduce model usefulness by rejecting sensitive but legitimate queries. We introduce Self-MOA (Self Multi-Objective Alignment), a fully automated framework for aligning small language models using weak supervision from automated evaluator models. Self-MOA operates as a closed loop that dynamically generates model-specific red team prompts, constructs preference data from model-generated responses, and aligns models via multi-objective preference optimization to jointly optimize for safety and helpfulness. Across multiple small language models and safety benchmarks, Self-MOA achieves a 12.41\% improvement in safety while preserving helpfulness, using as little as 11 times less training data than human-supervised alignment baselines. These results demonstrate that adaptive, automated alignment can reduce the dependence on static, human-curated safety pipelines in resource-constrained settings.

Seid Muhie Yimam, Daryna Dementieva, Tim Fischer et al.

Despite regulations imposed by nations and social media platforms, such as recent EU regulations targeting digital violence, abusive content persists as a significant challenge. Existing approaches primarily rely on binary solutions, such as outright blocking or banning, yet fail to address the complex nature of abusive speech. In this work, we propose a more comprehensive approach called Demarcation scoring abusive speech based on four aspect -- (i) severity scale; (ii) presence of a target; (iii) context scale; (iv) legal scale -- and suggesting more options of actions like detoxification, counter speech generation, blocking, or, as a final measure, human intervention. Through a thorough analysis of abusive speech regulations across diverse jurisdictions, platforms, and research papers we highlight the gap in preventing measures and advocate for tailored proactive steps to combat its multifaceted manifestations. Our work aims to inform future strategies for effectively addressing abusive speech online.

Somnath Banerjee, Maulindu Sarkar, Punyajoy Saha et al.

Recently, influence functions present an apparatus for achieving explainability for deep neural models by quantifying the perturbation of individual train instances that might impact a test prediction. Our objectives in this paper are twofold. First we incorporate influence functions as a feedback into the model to improve its performance. Second, in a dataset extension exercise, using influence functions to automatically identify data points that have been initially `silver' annotated by some existing method and need to be cross-checked (and corrected) by annotators to improve the model performance. To meet these objectives, in this paper, we introduce InfFeed, which uses influence functions to compute the influential instances for a target instance. Toward the first objective, we adjust the label of the target instance based on its influencer(s) label. In doing this, InfFeed outperforms the state-of-the-art baselines (including LLMs) by a maximum macro F1-score margin of almost 4% for hate speech classification, 3.5% for stance classification, and 3% for irony and 2% for sarcasm detection. Toward the second objective we show that manually re-annotating only those silver annotated data points in the extension set that have a negative influence can immensely improve the model performance bringing it very close to the scenario where all the data points in the extension set have gold labels. This allows for huge reduction of the number of data points that need to be manually annotated since out of the silver annotated extension dataset, the influence function scheme picks up ~1/1000 points that need manual correction.

Mithun Das, Rohit Raj, Punyajoy Saha et al.

Hate speech has become one of the most significant issues in modern society, having implications in both the online and the offline world. Due to this, hate speech research has recently gained a lot of traction. However, most of the work has primarily focused on text media with relatively little work on images and even lesser on videos. Thus, early stage automated video moderation techniques are needed to handle the videos that are being uploaded to keep the platform safe and healthy. With a view to detect and remove hateful content from the video sharing platforms, our work focuses on hate video detection using multi-modalities. To this end, we curate ~43 hours of videos from BitChute and manually annotate them as hate or non-hate, along with the frame spans which could explain the labelling decision. To collect the relevant videos we harnessed search keywords from hate lexicons. We observe various cues in images and audio of hateful videos. Further, we build deep learning multi-modal models to classify the hate videos and observe that using all the modalities of the videos improves the overall hate speech detection performance (accuracy=0.798, macro F1-score=0.790) by ~5.7% compared to the best uni-modal model in terms of macro F1 score. In summary, our work takes the first step toward understanding and modeling hateful videos on video hosting platforms such as BitChute.

Mithun Das, Somnath Banerjee, Punyajoy Saha

Online hatred is a growing concern on many social media platforms. To address this issue, different social media platforms have introduced moderation policies for such content. They also employ moderators who can check the posts violating moderation policies and take appropriate action. Academicians in the abusive language research domain also perform various studies to detect such content better. Although there is extensive research in abusive language detection in English, there is a lacuna in abusive language detection in low resource languages like Hindi, Urdu etc. In this FIRE 2021 shared task - "HASOC- Abusive and Threatening language detection in Urdu" the organizers propose an abusive language detection dataset in Urdu along with threatening language detection. In this paper, we explored several machine learning models such as XGboost, LGBM, m-BERT based models for abusive and threatening content detection in Urdu based on the shared task. We observed the Transformer model specifically trained on abusive language dataset in Arabic helps in getting the best performance. Our model came First for both abusive and threatening content detection with an F1scoreof 0.88 and 0.54, respectively.

Somnath Banerjee, Maulindu Sarkar, Nancy Agrawal et al.

Hate speech is considered to be one of the major issues currently plaguing online social media. Repeated and repetitive exposure to hate speech has been shown to create physiological effects on the target users. Thus, hate speech, in all its forms, should be addressed on these platforms in order to maintain good health. In this paper, we explored several Transformer based machine learning models for the detection of hate speech and offensive content in English and Indo-Aryan languages at FIRE 2021. We explore several models such as mBERT, XLMR-large, XLMR-base by team name "Super Mario". Our models came 2nd position in Code-Mixed Data set (Macro F1: 0.7107), 2nd position in Hindi two-class classification(Macro F1: 0.7797), 4th in English four-class category (Macro F1: 0.8006) and 12th in English two-class category (Macro F1: 0.6447).

Mithun Das, Punyajoy Saha, Ritam Dutt et al.

Hate speech is regarded as one of the crucial issues plaguing the online social media. The current literature on hate speech detection leverages primarily the textual content to find hateful posts and subsequently identify hateful users. However, this methodology disregards the social connections between users. In this paper, we run a detailed exploration of the problem space and investigate an array of models ranging from purely textual to graph based to finally semi-supervised techniques using Graph Neural Networks (GNN) that utilize both textual and graph-based features. We run exhaustive experiments on two datasets -- Gab, which is loosely moderated and Twitter, which is strictly moderated. Overall the AGNN model achieves 0.791 macro F1-score on the Gab dataset and 0.780 macro F1-score on the Twitter dataset using only 5% of the labeled instances, considerably outperforming all the other models including the fully supervised ones. We perform detailed error analysis on the best performing text and graph based models and observe that hateful users have unique network neighborhood signatures and the AGNN model benefits by paying attention to these signatures. This property, as we observe, also allows the model to generalize well across platforms in a zero-shot setting. Lastly, we utilize the best performing GNN model to analyze the evolution of hateful users and their targets over time in Gab.

Debjoy Saha, Naman Paharia, Debajit Chakraborty et al.

Social media often acts as breeding grounds for different forms of offensive content. For low resource languages like Tamil, the situation is more complex due to the poor performance of multilingual or language-specific models and lack of proper benchmark datasets. Based on this shared task, Offensive Language Identification in Dravidian Languages at EACL 2021, we present an exhaustive exploration of different transformer models, We also provide a genetic algorithm technique for ensembling different models. Our ensembled models trained separately for each language secured the first position in Tamil, the second position in Kannada, and the first position in Malayalam sub-tasks. The models and codes are provided.

Punyajoy Saha, Binny Mathew, Kiran Garimella et al.

WhatsApp is the most popular messaging app in the world. Due to its popularity, WhatsApp has become a powerful and cheap tool for political campaigning being widely used during the 2019 Indian general election, where it was used to connect to the voters on a large scale. Along with the campaigning, there have been reports that WhatsApp has also become a breeding ground for harmful speech against various protected groups and religious minorities. Many such messages attempt to instil fear among the population about a specific (minority) community. According to research on inter-group conflict, such `fear speech' messages could have a lasting impact and might lead to real offline violence. In this paper, we perform the first large scale study on fear speech across thousands of public WhatsApp groups discussing politics in India. We curate a new dataset and try to characterize fear speech from this dataset. We observe that users writing fear speech messages use various events and symbols to create the illusion of fear among the reader about a target community. We build models to classify fear speech and observe that current state-of-the-art NLP models do not perform well at this task. Fear speech messages tend to spread faster and could potentially go undetected by classifiers built to detect traditional toxic speech due to their low toxic nature. Finally, using a novel methodology to target users with Facebook ads, we conduct a survey among the users of these WhatsApp groups to understand the types of users who consume and share fear speech. We believe that this work opens up new research questions that are very different from tackling hate speech which the research community has been traditionally involved in.

Binny Mathew, Anurag Illendula, Punyajoy Saha et al.

With the ongoing debate on 'freedom of speech' vs. 'hate speech' there is an urgent need to carefully understand the consequences of the inevitable culmination of the two, i.e., 'freedom of hate speech' over time. An ideal scenario to understand this would be to observe the effects of hate speech in an (almost) unrestricted environment. Hence, we perform the first temporal analysis of hate speech on Gab.com, a social media site with very loose moderation policy. We first generate temporal snapshots of Gab from millions of posts and users. Using these temporal snapshots, we compute an activity vector based on DeGroot model to identify hateful users. The amount of hate speech in Gab is steadily increasing and the new users are becoming hateful at an increased and faster rate. Further, our analysis analysis reveals that the hate users are occupying the prominent positions in the Gab network. Also, the language used by the community as a whole seem to correlate more with that of the hateful users as compared to the non-hateful ones. We discuss how, many crucial design questions in CSCW open up from our work.