Sujan Dutta

Tharindu Cyril Weerasooriya, Sujan Dutta, Tharindu Ranasinghe et al.

Offensive speech detection is a key component of content moderation. However, what is offensive can be highly subjective. This paper investigates how machine and human moderators disagree on what is offensive when it comes to real-world social web political discourse. We show that (1) there is extensive disagreement among the moderators (humans and machines); and (2) human and large-language-model classifiers are unable to predict how other human raters will respond, based on their political leanings. For (1), we conduct a noise audit at an unprecedented scale that combines both machine and human responses. For (2), we introduce a first-of-its-kind dataset of vicarious offense. Our noise audit reveals that moderation outcomes vary wildly across different machine moderators. Our experiments with human moderators suggest that political leanings combined with sensitive issues affect both first-person and vicarious offense. The dataset is available through https://github.com/Homan-Lab/voiced.

Adel Khorramrouz, Sujan Dutta, Ashiqur R. KhudaBukhsh

In this paper, we present a computational analysis of the Persian language Twitter discourse with the aim to estimate the shift in stance toward gender equality following the death of Mahsa Amini in police custody. We present an ensemble active learning pipeline to train a stance classifier. Our novelty lies in the involvement of Iranian women in an active role as annotators in building this AI system. Our annotators not only provide labels, but they also suggest valuable keywords for more meaningful corpus creation as well as provide short example documents for a guided sampling step. Our analyses indicate that Mahsa Amini's death triggered polarized Persian language discourse where both fractions of negative and positive tweets toward gender equality increased. The increase in positive tweets was slightly greater than the increase in negative tweets. We also observe that with respect to account creation time, between the state-aligned Twitter accounts and pro-protest Twitter accounts, pro-protest accounts are more similar to baseline Persian Twitter activity.

Sujan Dutta, Deepak Pandita, Tharindu Cyril Weerasooriya et al.

Ensuring annotator quality in training and evaluation data is a key piece of machine learning in NLP. Tasks such as sentiment analysis and offensive speech detection are intrinsically subjective, creating a challenging scenario for traditional quality assessment approaches because it is hard to distinguish disagreement due to poor work from that due to differences of opinions between sincere annotators. With the goal of increasing diverse perspectives in annotation while ensuring consistency, we propose \texttt{ARTICLE}, an in-context learning (ICL) framework to estimate annotation quality through self-consistency. We evaluate this framework on two offensive speech datasets using multiple LLMs and compare its performance with traditional methods. Our findings indicate that \texttt{ARTICLE} can be used as a robust method for identifying reliable annotators, hence improving data quality.

Deepak Pandita, Tharindu Cyril Weerasooriya, Sujan Dutta et al.

Human feedback is essential for building human-centered AI systems across domains where disagreement is prevalent, such as AI safety, content moderation, or sentiment analysis. Many disagreements, particularly in politically charged settings, arise because raters have opposing values or beliefs. Vicarious annotation is a method for breaking down disagreement by asking raters how they think others would annotate the data. In this paper, we explore the use of vicarious annotation with analytical methods for moderating rater disagreement. We employ rater cohesion metrics to study the potential influence of political affiliations and demographic backgrounds on raters' perceptions of offense. Additionally, we utilize CrowdTruth's rater quality metrics, which consider the demographics of the raters, to score the raters and their annotations. We study how the rater quality metrics influence the in-group and cross-group rater cohesion across the personal and vicarious levels.

Sujan Dutta, Parth Srivastava, Vaishnavi Solunke et al.

Divorce is the legal dissolution of a marriage by a court. Since this is usually an unpleasant outcome of a marital union, each party may have reasons to call the decision to quit which is generally documented in detail in the court proceedings. Via a substantial corpus of 17,306 court proceedings, this paper investigates gender inequality through the lens of divorce court proceedings. While emerging data sources (e.g., public court records) on sensitive societal issues hold promise in aiding social science research, biases present in cutting-edge natural language processing (NLP) methods may interfere with or affect such studies. We thus require a thorough analysis of potential gaps and limitations present in extant NLP resources. In this paper, on the methodological side, we demonstrate that existing NLP resources required several non-trivial modifications to quantify societal inequalities. On the substantive side, we find that while a large number of court cases perhaps suggest changing norms in India where women are increasingly challenging patriarchy, AI-powered analyses of these court proceedings indicate striking gender inequality with women often subjected to domestic violence.

Somonnoy Banerjee, Sujan Dutta, Soumyajit Datta et al.

This paper makes three key contributions. First, via a substantial corpus of 51,278 interview questions sourced from 888 YouTube videos of mock interviews of Indian civil service candidates, we demonstrate stark gender bias in the broad nature of questions asked to male and female candidates. Second, our experiments with large language models show a strong presence of gender bias in explanations provided by the LLMs on the gender inference task. Finally, we present a novel dataset of 51,278 interview questions that can inform future social science studies.

Arka Dutta, Sujan Dutta, Rijul Magu et al.

Hallucinations pose a critical challenge to the real-world deployment of large language models (LLMs) in high-stakes domains. In this paper, we present a framework for stress testing factual fidelity in LLMs in the presence of adversarial nudge. Our framework consists of three steps. In the first step, we instruct the LLM to produce sets of truths and lies consistent with the closed domain in question. In the next step, we instruct the LLM to verify the same set of assertions as truths and lies consistent with the same closed domain. In the final step, we test the robustness of the LLM against the lies generated (and verified) by itself. Our extensive evaluation, conducted using five widely known proprietary LLMs across two closed domains of popular movies and novels, reveals a wide range of susceptibility to adversarial nudges: \texttt{Claude} exhibits strong resilience, \texttt{GPT} and \texttt{Grok} demonstrate moderate resilience, while \texttt{Gemini} and \texttt{DeepSeek} show weak resilience. Considering that a large population is increasingly using LLMs for information seeking, our findings raise alarm.

Arka Dutta, Adel Khorramrouz, Sujan Dutta et al.

This paper makes three contributions. First, it presents a generalizable, novel framework dubbed \textit{toxicity rabbit hole} that iteratively elicits toxic content from a wide suite of large language models. Spanning a set of 1,266 identity groups, we first conduct a bias audit of \texttt{PaLM 2} guardrails presenting key insights. Next, we report generalizability across several other models. Through the elicited toxic content, we present a broad analysis with a key emphasis on racism, antisemitism, misogyny, Islamophobia, homophobia, and transphobia. Finally, driven by concrete examples, we discuss potential ramifications.

Sujan Dutta, Sayantan Mahinder, Raviteja Anantha et al.

Reinforcement Learning from AI Feedback (RLAIF) has demonstrated significant potential across various domains, including mitigating harm in LLM outputs, enhancing text summarization, and mathematical reasoning. This paper introduces an RLAIF framework for improving the code generation abilities of lightweight (<1B parameters) LLMs. We specifically focus on code generation tasks that require writing appropriate API calls, which is challenging due to the well-known issue of hallucination in LLMs. Our framework extracts AI feedback from a larger LLM (e.g., GPT-3.5) through a specialized prompting strategy and uses this data to train a reward model towards better alignment from smaller LLMs. We run our experiments on the Gorilla dataset and meticulously assess the quality of the model-generated code across various metrics, including AST, ROUGE, and Code-BLEU, and develop a pipeline to compute its executability rate accurately. Our approach significantly enhances the fine-tuned LLM baseline's performance, achieving a 4.5% improvement in executability rate. Notably, a smaller LLM model (780M parameters) trained with RLAIF surpasses a much larger fine-tuned baseline with 7B parameters, achieving a 1.0% higher code executability rate.