CS Paper Radar

Veterans' Mental Health Crisis Severity Annotation: Do LLMs Agree More Than Human Experts?

K M Sajjadul Islam, Pranav Yaramalla, Nadiyah Johnson, Mark Flower, Praveen Madiraju, Zeno Franco

cs.AI ·submitted 2026-11-05 22:18 UTC · v2
publishedcode/data

Abstract

Large language models (LLMs) offer a scalable approach for annotating mental health and crisis-related social media content, but their reliability in safety-critical tasks remains unclear. This study evaluates LLM-based crisis-severity annotation of veteran-related social media posts against annotations from eight Veteran Subject Matter Experts (SMEs). Using 149 tweets and five severity categories, we evaluate 15 open-source and 10 proprietary LLMs across eight prompt variants, examining annotation performance, inter-model agreement, prompt sensitivity, and high-severity crisis detection. Human reviewers showed higher ordinal than nominal agreement (α = 0.494), indicating that disagreements often occurred between neighboring severity levels. In contrast, selected LLMs demonstrated substantially higher inter-model consistency, with ordinal agreement (α) reaching 0.8336 under the baseline prompt. Proprietary models generally showed better agreement with expert annotations and greater stability than open-source models. GPT-5.6-Terra achieved the best overall performance, while the open-source Qwen2.5-7B showed the strongest performance in identifying Acute Crisis with Suicidal Ideation cases. However, higher model consistency did not necessarily correspond to agreement with human judgments or reliable detection of rare high-severity cases. Label-flipped demonstration experiments also revealed model-specific sensitivity to misleading in-context examples, while explanatory demonstrations did not consistently improve performance. Overall, LLMs show promise as scalable complements to expert annotation, but require careful evaluation before use in high-stakes mental health applications.

Comment
Accepted in IEEE Big Data 2026

Closest papers

Nearest neighbours by meaning across the whole index.

Edge-Deployable LLMs for Classification and Value-of-Information Ranking of Disaster-Related Tweets

cs.IR ·2026-11-06 ·Christopher D. Molek, K. Brent Venable, Niranjan Suri, Stephen Russell
similarity 0.72
Abstract · 6 pages, 6 figures, 1 table, IEEE MILCOM 2026

Humanitarian-aid and disaster relief decisions often rely on noisy, incomplete, and uncertain information from social media. Although substantial research has focused on automating the classification of such data, limited work has examined whether individual posts contain sufficient information to support decisions to act. In this study, we use large language models (LLMs) to (1) classify social media content and (2) assess its Value-of-Information (VoI) in terms of operational actionability. We evaluate seven LLMs for the classification of 77,484 tweets across 10 humanitarian-aid categories and identify the Qwen3 (8B parameters) model as the top-performing classifier, achieving the highest F1 scores in all but one category. In a second task, the classified categories serve as contextual inputs for assessing VoI with respect to responder actionability across six well established metrics for each tweet. We observed that the Gemma 3 (4B parameters) model outperformed the other models on this task, achieving the highest recall across the majority of categories. This work demonstrates that relatively small, edge-deployable LLMs can prioritize actionable information in resource-constrained environments. By integrating VoI assessment, our framework advances the automated elicitation of high-utility knowledge from unstructured social media streams for disaster-response decision-making and highlights a pathway toward tactical edge applications using LLMs that fit on a single 8 GB GPU.

0.22

CrisisFake: Benchmark Validity of AI-Generated Text Detection for Disaster Social Sensing

cs.CL cs.CY ·2026-09-21 ·Xiaoshan Zhou, Zaifu Zhan
benchmark similarity 0.68
Abstract

Disaster social sensing converts public social-media posts into evidence for situational awareness and humanitarian response, but plausible LLM-generated posts can contaminate this information stream and distort assessments of needs, damage, and resource priorities. This study empirically investigates whether text-based detectors can distinguish human-authored from LLM-generated disaster posts and what textual cues underlie their judgments. We construct CrisisFake, a Qwen2.5-7B-based dataset of 12,000 texts organized into 3,000 matched semantic units from nine disasters. Each unit contains an original human post, a minimally LLM-proofread human post, a fact-preserving LLM-generated post produced using LoRA, and an affectively reframed version of the artificial post. A separate 6,000-text corpus spanning 42 disaster events is independently constructed to support model selection and threshold calibration. We evaluate OSM-Det, Fast-DetectGPT, Binoculars, and direct LLM judges across five model families, and further examine how disaster-domain calibration and superficial linguistic cues, such as retweet markers, user mentions, hashtags, URLs, punctuation, and text length, affect detector performance. Across fourteen frozen configurations, AUROC ranges from 0.402 to 0.517, while the best prospective recall at a calibration-derived low-false-positive operating point is only 3.6%, indicating near-chance discrimination. For OSM-Det, a disaster-calibrated linear head improves AUROC to 0.817; however, a seven-feature textual classifier alone reaches AUROC 0.784 on the original-versus-factual-LLM contrast, and neutralizing identified surface asymmetries reduces the corresponding linear-head AUROC from 0.733 to 0.594. These findings provide empirical evidence that LLM-generated text detection is largely driven by linguistic cues and remains insufficiently robust for short-form disaster social media.

0.05

Lightweight Machine Learning for Classifying Disaster-Related Social Media Posts

cs.NI cs.LG ·2026-10-17 ·Chengjun Jiang, Milena Radenkovic
similarity 0.67
Abstract

Classifying disaster-related social-media posts into humanitarian information labels is challenging because individual messages are typically short and context-limited, humanitarian categories can overlap semantically, and the CrisisBench label distribution is strongly imbalanced. This study investigates whether lightweight lexical machine-learning methods can provide effective and computationally efficient classification under these conditions. Using the CrisisBench v1.0 English humanitarian task, comprising 87,455 records across predefined Train, Development and Test splits, we conduct a controlled comparison of TF-IDF representations and lightweight classifiers, with Macro-F1 as the primary model-selection metric. Beyond aggregate predictive performance, the study jointly examines per-label behaviour, directional error patterns, relative decision margins, computational requirements, and the effect of domain-specific training through a tropical-cyclone specialist model. The selected balanced LinearSVC with unigram-bigram TF-IDF and C=0.5 achieved an Accuracy of 0.8282, a Macro-F1 of 0.6578 and a Weighted-F1 of 0.8315 on the held-out Test split. Lower decision margins were substantially associated with prediction errors, while the tropical-cyclone specialist achieved modestly higher in-domain aggregate performance with substantially lower feature and fitting requirements. The study contributes a controlled, label-sensitive and resource-aware evaluation of lightweight humanitarian text classification that characterises not only aggregate performance, but also where these models fail and how their behaviour changes under domain-specific training.

0.23

DistressBench: Evaluating Multi-Dimensional Crisis Support in Large Language Models

cs.CL cs.CY ·2026-10-08 ·Drew Rein, Patrick Oathout, Vishal Kumar, Udari Madhushani Sehwag
similarity 0.66
Abstract · 39 pages, 9 figures, 13 tables

Language models have been adopted faster than evaluation practices that reliably assess their risks, and a growing share of use is affective rather than informational. Within that traffic users disclose suicide and self-harm (SSH) crises. Existing safety benchmarks score models primarily on refusal, but refusal is not the desired behavior here; the conversation requires a substantive supportive response. We introduce DistressBench, 718 English conversations with clinician-authored user turns spanning 23 SSH subcategories, each paired with a task-specific rubric of weighted atomic criteria across seven dimensions of crisis support and a clinician reference response. A model's score is the weighted average over the criteria its reply satisfies, reading directly as the fraction of required care delivered. The best of 25 frontier models satisfies 88.4% of weighted criteria and the median model 71.4%, against 98.7% for clinician references scored on the same criteria. Pooled across all 25 models, in 35.3% of conversations where a model satisfied every criterion for recognizing the crisis, it failed to route the user toward human help. The shortfall concentrates in the dimensions requiring the model to act rather than withhold. Contrasting each dimension's median model against its best separates capability limits from choices: de-escalation reaches 74.6% in the strongest model against a median of 46.0%, and safety disclaimers 84.6% against 8.3%. Multi-turn conversations score below single-turn ones (68.5% against 74.2%). We release the corpus, rubrics, judge calibration labels, and a companion annex of 20 long-horizon red-team transcripts. These results argue for evaluating crisis safety on the care a model delivers rather than on what it declines to say, and for treating de-escalation and resource routing as deployment requirements.

0.21

Building Trustworthy Mental Health Benchmarks on Bluesky: A Validation-Aware Weak-Supervision Framework

cs.AI ·2026-09-19 ·Gaurab Chhetri, Anandi Dutta, Subasish Das
similarity 0.65
Abstract · This is the author's preprint version of a paper accepted for presentation at HICSS 60 (Hawaii International Conference on System Sciences),…

Decentralized social media platforms create new opportunities and challenges for computational mental health research because data access, moderation, labeling, and deployment responsibilities are distributed across multiple technical and governance layers. This paper presents a validation-aware weak-supervision system for constructing and evaluating suicidal ideation (SI) and broader mental health (MH) disclosure benchmarks on Bluesky, a decentralized social media platform built on the AT Protocol. The system integrates public firehose collection, task-specific lexicon filtering, Llama-3-8B-assisted binary annotation, human-adjudicated validation subsets, and transformer-based model benchmarking. Using this pipeline, we construct two task-specific corpora containing 8,346 SI-labeled posts and 9,988 MH-labeled posts. The evaluation shows that model performance depends strongly on both task definition and validation protocol. BERT+LSTM achieves the highest SI stratified cross-validation F1-score, RoBERTa achieves the strongest SI holdout F1-score, and DistilRoBERTa achieves the best MH cross-validation F1-score. Human validation reveals different weak-label failure modes across tasks, with SI labels dominated by false negatives and MH labels dominated by false positives. These findings show that decentralized social media can support reproducible mental health benchmarking, but only when system design, label provenance, validation strategy, and deployment constraints are evaluated together.

0.46

Can LLMs Follow the Pulse of a Crisis? Evaluating Crisis Sentiment in Bangladesh's July Uprising

cs.CL ·2026-09-15 ·Md. Samiul Alim, Mahir Shahriar Tamim, Tanvir Ahmed Khan, Sharjil Khan, Rafia Ferdous Duti +2 more
publishedbenchmark similarity 0.65
Abstract · Accepted at AACL

Crisis sentiment analysis is especially challenging for low-resource languages such as Bangla, where language, context, and public reaction shift rapidly. We introduce UNRESTSENT200K, a Bangla crisis sentiment dataset with approximately 200K Facebook and YouTube comments from the July-August 2024 Bangladesh uprising. The dataset covers five event-aligned phases, from early escalation and internet blackout to regime transition and a later flood crisis. Each comment is linked to its parent post, enabling evaluation with and without discourse context. All comments are annotated through a fully human process involving 14 native Bangla-speaking annotators and senior validation, achieving substantial agreement (kappa = 0.73, alpha = 0.71) and 94.2% blind-audit agreement. We benchmark fine-tuned encoders, prompted LLMs, and LoRA-tuned LLMs. Results show that parent-post context consistently improves performance, while temporal shift across phases causes large performance drops. Strong LLMs perform well, but still struggle with sarcasm, implicit political references, and phase-dependent meaning. UNRESTSENT200K provides a benchmark for studying context-aware and temporally robust sentiment analysis in low-resource crisis discourse. UNRESTSENT200K is available at https://sami0055.github.io/UNRESTSENT200K/

0.37

Nobody Truly Agrees on Sentiment: Humans, Bespoke Tools, and LLMs Struggle with Social Media Texts

cs.CL ·2026-10-07 ·Himarsha R. Jayanetti, Sivakanesan Dhanushkanda, Shuai Hao, Michael L. Nelson, Michele C. Weigle
published similarity 0.64
Abstract · 11 pages, 1 figure, 2 tables, accepted for publication at TPDL 2026

Social media is a rich source of real-time public sentiment, but widely used sentiment analysis tools are often applied without understanding their limitations. In this study, we evaluate the inter-rater reliability of three bespoke sentiment analysis tools (TextBlob, VADER, and Twitter-roBERTa-base) and three large language models (LLMs: Qwen3-32B, GPT-OSS-120B, Llama-4-Maverick-17B) against six human raters across 100 tweets. We measured agreement using two statistical measures: Cohen's kappa for pairwise comparisons and Fleiss' kappa for multiple raters. Even among the human raters, our results showed only fair agreement, highlighting the subjectivity of sentiment analysis. Higher agreement was observed under the binary sentiment classification (negative vs. non-negative and positive vs. non-positive) than under the three-class classification across both humans and automated tools. The Twitter-roBERTa-base model showed the strongest alignment with human ratings, outperforming both bespoke sentiment tools and LLMs, particularly in distinguishing negative versus non-negative sentiment. LLMs showed substantial agreement among themselves and moderate to substantial alignment with humans, performing better in positive vs. non-positive classifications. Our findings underscore that domain-specific fine-tuning remains crucial for reliable social media sentiment analysis, and human-centered evaluation remains essential for establishing gold-standard labels.

0.32

Whose Assessment of Distress? Community Perspectives and LLM Alignment on Well-Being Posts

cs.CL ·2026-08-29 ·Andrew Aquilina, Xiang Lorraine Li, Yu-Ru Li
similarity 0.64
Abstract

Judgments about psychological distress are socially situated: what counts as concerning hinges on community norms around emotional expression, vulnerability, and help-seeking. Yet large language models (LLMs) used for distress detection are typically aligned to a single, undifferentiated standard. How well do these models capture the perspectives of the communities whose language they assess? We address this question through a perspectivist annotation study in which 321 participants provided 9,587 judgments on 1,198 Reddit posts spanning six identity-based communities, yielding community-specific labels. Raters in the contextualized in-group condition show a modest tendency to agree more with their community than uncontextualized out-group raters (OR = 1.18), an effect varying significantly across communities. We then evaluate nine open-weight LLM configurations and four frontier configurations against these labels. Open-weight LLMs systematically over-estimate distress: when communities perceive none-to-mild distress, these models achieve only 31-44% accuracy, predominantly producing false positives. GPT-5 and Gemini 2.5 Pro show the same none-to-mild inflation even when their full-sample over/under rates are mixed, while Claude Opus 4 is more conservative. This pattern does not simply mirror an outsider reading position: uncontextualized out-group human aggregates were nearly symmetric, with 18% over-estimation versus 19% under-estimation. Instead, the models that inflate none-to-mild cases exhibit a distress prior that exceeds both contextualized in-group and uncontextualized out-group human judgments. These findings have implications for equitable AI deployment in mental health contexts, where miscalibrated distress detection may unevenly affect the communities being assessed.

0.24