Aggregation Artifacts in Subjective Tasks Collapse Large Language Models' Posteriors

In-context Learning (ICL) has become the primary method for performing natural language tasks with Large Language Models (LLMs). The knowledge acquired during pre-training is crucial for this few-shot capability, providing the model with task priors. However, recent studies have shown that ICL predominantly relies on retrieving task priors rather than "learning" to perform tasks. This limitation is particularly evident in complex subjective domains such as emotion and morality, where priors significantly influence posterior predictions. In this work, we examine whether this is the result of the aggregation used in corresponding datasets, where trying to combine low-agreement, disparate annotations might lead to annotation artifacts that create detrimental noise in the prompt. Moreover, we evaluate the posterior bias towards certain annotators by grounding our study in appropriate, quantitative measures of LLM priors. Our results indicate that aggregation is a confounding factor in the modeling of subjective tasks, and advocate focusing on modeling individuals instead. However, aggregation does not explain the entire gap between ICL and the state of the art, meaning other factors in such tasks also account for the observed phenomena. Finally, by rigorously studying annotator-level labels, we find that it is possible for minority annotators to both better align with LLMs and have their perspectives further amplified.

Modeling Information Narrative Detection and Evolution on Telegram during the Russia-Ukraine War

Following the Russian Federation's full-scale invasion of Ukraine in February 2022, a multitude of information narratives emerged within both pro-Russian and pro-Ukrainian communities online. As the conflict progresses, so too do the information narratives, constantly adapting and influencing local and global community perceptions and attitudes. This dynamic nature of the evolving information environment (IE) underscores a critical need to fully discern how narratives evolve and affect online communities. Existing research, however, often fails to capture information narrative evolution, overlooking both the fluid nature of narratives and the internal mechanisms that drive their evolution. Recognizing this, we introduce a novel approach designed to both model narrative evolution and uncover the underlying mechanisms driving them. In this work we perform a comparative discourse analysis across communities on Telegram covering the initial three months following the invasion. First, we uncover substantial disparities in narratives and perceptions between pro-Russian and pro-Ukrainian communities. Then, we probe deeper into prevalent narratives of each group, identifying key themes and examining the underlying mechanisms fueling their evolution. Finally, we explore influences and factors that may shape the development and spread of narratives.

Larger Language Models Don't Care How You Think: Why Chain-of-Thought Prompting Fails in Subjective Tasks

In-Context Learning (ICL) in Large Language Models (LLM) has emerged as the dominant technique for performing natural language tasks, as it does not require updating the model parameters with gradient-based methods. ICL promises to "adapt" the LLM to perform the present task at a competitive or state-of-the-art level at a fraction of the computational cost. ICL can be augmented by incorporating the reasoning process to arrive at the final label explicitly in the prompt, a technique called Chain-of-Thought (CoT) prompting. However, recent work has found that ICL relies mostly on the retrieval of task priors and less so on "learning" to perform tasks, especially for complex subjective domains like emotion and morality, where priors ossify posterior predictions. In this work, we examine whether "enabling" reasoning also creates the same behavior in LLMs, wherein the format of CoT retrieves reasoning priors that remain relatively unchanged despite the evidence in the prompt. We find that, surprisingly, CoT indeed suffers from the same posterior collapse as ICL for larger language models. Code is avalaible at https://github.com/gchochla/cot-priors.

Towards Safer Online Spaces: Simulating and Assessing Intervention Strategies for Eating Disorder Discussions

Eating disorders are complex mental health conditions that affect millions of people around the world. Effective interventions on social media platforms are crucial, yet testing strategies in situ can be risky. We present a novel LLM-driven experimental testbed for simulating and assessing intervention strategies in ED-related discussions. Our framework generates synthetic conversations across multiple platforms, models, and ED-related topics, allowing for controlled experimentation with diverse intervention approaches. We analyze the impact of various intervention strategies on conversation dynamics across four dimensions: intervention type, generative model, social media platform, and ED-related community/topic. We employ cognitive domain analysis metrics, including sentiment, emotions, etc., to evaluate the effectiveness of interventions. Our findings reveal that civility-focused interventions consistently improve positive sentiment and emotional tone across all dimensions, while insight-resetting approaches tend to increase negative emotions. We also uncover significant biases in LLM-generated conversations, with cognitive metrics varying notably between models (Claude-3 Haiku $>$ Mistral $>$ GPT-3.5-turbo $>$ LLaMA3) and even between versions of the same model. These variations highlight the importance of model selection in simulating realistic discussions related to ED. Our work provides valuable information on the complex dynamics of ED-related discussions and the effectiveness of various intervention strategies.

Improving and Assessing the Fidelity of Large Language Models Alignment to Online Communities

Large language models (LLMs) have shown promise in representing individuals and communities, offering new ways to study complex social dynamics. However, effectively aligning LLMs with specific human groups and systematically assessing the fidelity of the alignment remains a challenge. This paper presents a robust framework for aligning LLMs with online communities via instruction-tuning and comprehensively evaluating alignment across various aspects of language, including authenticity, emotional tone, toxicity, and harm. We demonstrate the utility of our approach by applying it to online communities centered on dieting and body image. We administer an eating disorder psychometric test to the aligned LLMs to reveal unhealthy beliefs and successfully differentiate communities with varying levels of eating disorder risk. Our results highlight the potential of LLMs in automated moderation and broader applications in public health and social science research.

Feelings about Bodies: Emotions on Diet and Fitness Forums Reveal Gendered Stereotypes and Body Image Concerns

The gendered expectations about ideal body types can lead to body image concerns, dissatisfaction, and in extreme cases, disordered eating and other psychopathologies across the gender spectrum. While research has focused on pro-anorexia online communities that glorify the 'thin ideal', less attention has been given to the broader spectrum of body image concerns or how emerging disorders like muscle dysmorphia ('bigorexia') present in online discussions. To address these gaps, we analyze 46 Reddit discussion forums related to diet, fitness, and associated mental health challenges. Using membership structure analysis and transformer-based language models, we project these communities along gender and body ideal axes, revealing complex interactions between gender, body ideals, and emotional expression. Our findings show that feminine-oriented communities generally express more negative emotions, particularly in thinness-promoting forums. Conversely, communities focused on the muscular ideal exhibit less negativity, regardless of gender orientation. We also uncover a gendered pattern in emotional indicators of mental health challenges, with communities discussing serious issues aligning more closely with thinness-oriented, predominantly feminine-leaning communities. By revealing the gendered emotional dynamics of online communities, our findings can inform the development of more effective content moderation approaches that facilitate supportive interactions, while minimizing exposure to potentially harmful content.

COMMUNITY-CROSS-INSTRUCT: Unsupervised Instruction Generation for Aligning Large Language Models to Online Communities

Social scientists use surveys to probe the opinions and beliefs of populations, but these methods are slow, costly, and prone to biases. Recent advances in large language models (LLMs) enable creating computational representations or "digital twins" of populations that generate human-like responses mimicking the population's language, styles, and attitudes. We introduce Community-Cross-Instruct, an unsupervised framework for aligning LLMs to online communities to elicit their beliefs. Given a corpus of a community's online discussions, Community-Cross-Instruct automatically generates instruction-output pairs by an advanced LLM to (1) finetune an foundational LLM to faithfully represent that community, and (2) evaluate the alignment of the finetuned model to the community. We demonstrate the method's utility in accurately representing political and fitness communities on Reddit. Unlike prior methods requiring human-authored instructions, Community-Cross-Instruct generates instructions in a fully unsupervised manner, enhancing scalability and generalization across domains. This work enables cost-effective and automated surveying of diverse online communities.

#EpiTwitter: Public Health Messaging During the COVID-19 Pandemic

Effective communication during health crises is critical, with social media serving as a key platform for public health experts (PHEs) to engage with the public. However, it also amplifies pseudo-experts promoting contrarian views. Despite its importance, the role of emotional and moral language in PHEs' communication during COVID-19 remains under explored. This study examines how PHEs and pseudo-experts communicated on Twitter during the pandemic, focusing on emotional and moral language and their engagement with political elites. Analyzing tweets from 489 PHEs and 356 pseudo-experts from January 2020 to January 2021, alongside public responses, we identified key priorities and differences in messaging strategy. PHEs prioritize masking, healthcare, education, and vaccines, using positive emotional language like optimism. In contrast, pseudo-experts discuss therapeutics and lockdowns more frequently, employing negative emotions like pessimism and disgust. Negative emotional and moral language tends to drive engagement, but positive language from PHEs fosters positivity in public responses. PHEs exhibit liberal partisanship, expressing more positivity towards liberals and negativity towards conservative elites, while pseudo-experts show conservative partisanship. These findings shed light on the polarization of COVID-19 discourse and underscore the importance of strategic use of emotional and moral language by experts to mitigate polarization and enhance public trust.

Trust and Terror: Hazards in Text Reveal Negatively Biased Credulity and Partisan Negativity Bias

Socio-linguistic indicators of text, such as emotion or sentiment, are often extracted using neural networks in order to better understand features of social media. One indicator that is often overlooked, however, is the presence of hazards within text. Recent psychological research suggests that statements about hazards are more believable than statements about benefits (a property known as negatively biased credulity), and that political liberals and conservatives differ in how often they share hazards. Here, we develop a new model to detect information concerning hazards, trained on a new collection of annotated X posts, as well as urban legends annotated in previous work. We show that not only does this model perform well (outperforming, e.g., zero-shot human annotator proxies, such as GPT-4) but that the hazard information it extracts is not strongly correlated with other indicators, namely moral outrage, sentiment, emotions, and threat words. (That said, consonant with expectations, hazard information does correlate positively with such emotions as fear, and negatively with emotions like joy.) We then apply this model to three datasets: X posts about COVID-19, X posts about the 2023 Hamas-Israel war, and a new expanded collection of urban legends. From these data, we uncover words associated with hazards unique to each dataset as well as differences in this language between groups of users, such as conservatives and liberals, which informs what these groups perceive as hazards. We further show that information about hazards peaks in frequency after major hazard events, and therefore acts as an automated indicator of such events. Finally, we find that information about hazards is especially prevalent in urban legends, which is consistent with previous work that finds that reports of hazards are more likely to be both believed and transmitted.

Harmful Speech Detection by Language Models Exhibits Gender-Queer Dialect Bias

Content moderation on social media platforms shapes the dynamics of online discourse, influencing whose voices are amplified and whose are suppressed. Recent studies have raised concerns about the fairness of content moderation practices, particularly for aggressively flagging posts from transgender and non-binary individuals as toxic. In this study, we investigate the presence of bias in harmful speech classification of gender-queer dialect online, focusing specifically on the treatment of reclaimed slurs. We introduce a novel dataset, QueerReclaimLex, based on 109 curated templates exemplifying non-derogatory uses of LGBTQ+ slurs. Dataset instances are scored by gender-queer annotators for potential harm depending on additional context about speaker identity. We systematically evaluate the performance of five off-the-shelf language models in assessing the harm of these texts and explore the effectiveness of chain-of-thought prompting to teach large language models (LLMs) to leverage author identity context. We reveal a tendency for these models to inaccurately flag texts authored by gender-queer individuals as harmful. Strikingly, across all LLMs the performance is poorest for texts that show signs of being written by individuals targeted by the featured slur (F1 <= 0.24). We highlight an urgent need for fairness and inclusivity in content moderation systems. By uncovering these biases, this work aims to inform the development of more equitable content moderation practices and contribute to the creation of inclusive online spaces for all users.