Pith. sign in

REVIEW 4 major objections 7 minor 39 references

Just a Scratch: Enhancing LLM Capabilities for Self-harm Detection through Intent Differentiation and Emoji Interpretation

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Fine-tuning LLMs with curated emoji interpretations lifts self-harm detection F1 to 0.88

desk verdict Useful new dataset and sensible multitask framework, but the CESM-100 evaluation looks like corpus leakage and the rationale metrics are near-tautological. read the letter →

arxiv 2506.05073 v1 pith:EKF7RO4I submitted 2025-06-05 cs.CL

Soumitra Ghosh , Gopendra Vikram Singh , Shambhavi , Sabarna Choudhury , Asif Ekbal This is my paper · ORCID
classification cs.CL
keywords self-harmdetectionemojiinterpretationlargelanguagemodelsmultitaskfine-tuningspanextractioncasualmentionseriousintentCESM-100
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that large language models misread self-harm on social media because they flatten two distinct intents—casual mentions (hyperbole, dark humor) and serious intent (genuine distress)—and ignore what emojis do to a post's meaning. To fix this, the authors build CESM-100, a hand-curated matrix of 100 emojis with self-harm-related contextual interpretations, and SHINES, a 5,206-post dataset annotated for self-harm labels, casual-mention spans, serious-intent spans, and emoji context. They then fine-tune Llama 3, Mental-Alpaca, and MentalLlama in a multitask setup that predicts the label, extracts both span types, and generates an explanation. The headline result is that this intent- and emoji-aware multitask fine-tuning raises Llama 3's self-harm classification F1 to 0.88, from 0.83 for single-task fine-tuning without CESM-100 and 0.84 for multitask without CESM-100. If the gain is real, injecting a small curated emoji resource into LLM training is a cheap way to make detection and explanation more accurate without larger models.

What carries the argument

The load-bearing machinery is the Centennial Emoji Sensitivity Matrix (CESM-100): a curated table of 100 emojis, each paired with its usual meaning, a self-harm-specific contextual meaning, and two three-level ratings—casual-mention chance and serious-intent chance. Fine-tuning prompts insert the post text plus the CESM-100 interpretations of any emojis it contains, and the multitask objective trains the model simultaneously to output the binary self-harm label, the spans for casual mentions, the spans for serious intents, and a rationale that references those spans and emoji meanings. The span-extraction tasks are what force the model to differentiate intent explicitly; CESM-100 is what tells it what an emoji is doing in a self-harm context.

What would settle it

Rebuild CESM-100 using only the 80% training split (and, symmetrically, re-run the emoji-frequency analysis within splits), then repeat the multitask fine-tuning on the held-out test set; if the F1 advantage over the no-CESM-100 baseline collapses or reverses, the 0.88 result was leakage. A simpler check: scan the SHINES test posts for emojis whose CESM-100 interpretation cites an attribute that only appears in test-set data.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that the ambiguity in self-harm language reduces to an intent-differentiation problem, and that emojis carry much of the information needed to resolve it. The authors construct CESM-100, in which each of 100 emojis gets a contextual self-harm meaning plus separate casual-mention-chance and serious-intent-chance ratings, and SHINES, a corpus of 5,206 posts with self-harm labels and CM/SI span annotations. They fine-tune three LLMs with a multitask loss that jointly optimizes self-harm classification, casual-mention span extraction, and serious-intent span extraction, feeding CESM-100 interpretations into the prompt. Their results show the largest absolute gain from adding CESM-100 to multitask fine-tuning: Llama 3 reaches a self-harm F1 of 0.88, and the same combination improves span-extraction F1 and the semantic similarity of generated rationales. The paper therefore claims that intent differentiation, operationalized as span labels, plus emoji interpretation, operationalized as CESM-100, is what lets an LLM separate 'just a scratch' from a real cry for help.

Load-bearing premise

The load-bearing premise is that CESM-100's emoji interpretations were derived from the training distribution only; if the test posts' emojis influenced those interpretations, the reported 4-point F1 gain from CESM-100 would measure information leakage rather than a reusable resource.

Editorial extensions

If this is right

  • Multitask fine-tuning with CESM-100 outscored single-task fine-tuning by 3 F1 points for Llama 3 (0.88 vs 0.83) and by 2 points for Mental-Alpaca and MentalLlama, so the span-extraction auxiliary tasks generalize to the primary classification task.
  • Removing CESM-100 from multitask fine-tuning drops Llama 3's F1 from 0.88 to 0.84, and Mental-Alpaca's CM-span F1 from 0.83 to 0.81, indicating the emoji resource, not just the multitask loss, carries part of the gain.
  • Rationale generation improves along with detection: Llama 3's rationale semantic similarity rises from 0.85 without CESM-100 to 0.88 with it, so the same framework doubles as an explainability tool.
  • Introducing emoji noise (repositioning or random replacement) in 20% of training posts reduces self-harm classification F1 by 2.5%, evidence that emoji context is a causal ingredient, not a confound, in the model's decisions.
  • The framework transfers across one open-domain model (Llama 3) and two mental-health-tuned models (Mental-Alpaca, MentalLlama), with F1 gains in the same direction for all three.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: Because CESM-100 was built from emojis extracted from the collected posts and no train/test separation is reported for that extraction, a decisive test is to rebuild CESM-100 from the training split only; if the F1 gain over the no-CESM-100 baseline mostly disappears, the reported 0.88 reflects test-set leakage rather than a reusable resource.
  • Editorial inference: The paper analyzes emoji frequency on the full dataset before splitting, which risks letting aggregate statistics guide annotation; re-running the analysis within the training split only would show whether the CM/SI ratings generalize.
  • Editorial inference: The authors' own limitation notes say the full SHINES dataset is 'to be released pending research approval,' which contradicts the abstract's 'publicly available' claim; resolving that discrepancy matters for anyone trying to reproduce the 0.88 result.
  • Editorial inference: A natural extension the authors do not test is whether CESM-100 helps on newer or multimodal models or on non-Reddit platforms, where emoji norms and sarcasm conventions differ; the paper's own platform-bias discussion suggests this boundary is untested.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper introduces CESM-100, a curated resource of 100 emojis with contextual self-harm interpretations, and SHINES, a dataset of 5,206 Reddit posts annotated for self-harm/non-self-harm, casual-mention (CM) spans, and serious-intent (SI) spans. The authors fine-tune three LLMs (Llama 3.1-8B-Instruct, Mental-Alpaca-7B, MentaLLaMA-chat-7B) in zero-shot, few-shot, single-task, and multitask settings, augmenting inputs with CESM-100 emoji interpretations and generating rationales for classification decisions. The headline result (Table 9) is that multitask fine-tuning with CESM-100 raises Llama 3's self-harm classification F1 from 0.84 (multitask without CESM-100) to 0.88, with parallel gains in span extraction and rationale-generation metrics. The paper claims that coupling intent differentiation (CM vs. SI) with contextual emoji interpretation enhances both detection and explanation performance.

Significance. If valid, the contributions are useful for mental-health NLP: SHINES is a novel annotated resource pairing binary self-harm labels with CM/SI span labels and emoji interpretations, and CESM-100 is a concrete operationalization of emoji context for a safety-relevant task. The paper deserves credit for evaluating three models, for ablations that separate the multitask and CESM-100 contributions (Ablations 1 and 2 in Table 9), for attempting paired significance tests, for a noise-sensitivity analysis of emoji perturbations, for using synthetic posts only in training, and for a detailed ethical and limitations section. The central empirical claims, however, are not yet independently verified: the CESM-100 gain is threatened by possible train/test information leakage in the resource's construction, and the rationale-generation metrics reward copying of gold spans rather than measuring explanation quality. If these two concerns are resolved with re-run experiments, the paper would be a solid contribution; as it stands, the main quantitative claims rest on circular evaluation.

major comments (4)
  1. [§4.1, §5, §7; Table 9] Section 4.1 states that CESM-100 was built from "a subset of emojis ... extracted from the collected posts," and the emoji-frequency analyses of Section 5 and Appendices A.4-A.5 are computed over the full 5,206-post SHINES collection, while the 80/20 train/test split is introduced only in Section 7 with no documented separation during CESM-100 curation. Because CESM-100's CM/SI chance ratings are plausibly informed by observed emoji-label co-occurrence over the whole corpus, the test-time input augmentation gives the model emoji interpretations that may encode test-set label statistics. The headline result that adding CESM-100 raises Llama 3's multitask SHC F1 from 0.84 to 0.88 (Table 9) is therefore susceptible to information leakage: the paired t-tests in Section 8 compare the model on the same leaked resource and cannot distinguish a reusable emoji resource from memorized corpus-level associations. Please rebuild CESM-100 using only training-split data (or an independent emoji-usage corpus) and re-run the affected experiments and statistical comparisons.
  2. [§A.7; Table 12] The rationale-generation metrics defined in Section A.7 make the RG results near-tautological: Relevance "checks if all spans are present in the rationale," and Coherence and Semantic Similarity compute cosine similarity between the rationale and the combined CM/SI span text. But the rationale-generation prompt in Table 12 already injects the gold classification, gold CM/SI spans, and CESM-100 emoji interpretations into the input, so the model need only copy the provided spans to score highly. The reported RG gains (e.g., Semantic Similarity 0.88 with CESM-100 versus 0.85 without, Table 9) therefore measure span copying rather than explanation quality, and they cannot support the abstract's claim that the framework improves "explanation tasks." The RG protocol should be re-run with model-predicted spans (not gold spans) in the input, and ideally complemented by human evaluation of rationale faithfulness and informativeness.
  3. [Table 2; §3.2] Table 2 reports that 2,707 of 2,707 non-self-harm posts contain CM spans, 2,488 of 2,499 self-harm posts contain SI spans, and only 34 self-harm posts contain CM spans. The CM/SI span labels are thus almost perfectly determined by the binary self-harm label, so the multitask benefit attributed to "intent differentiation" may simply be token-level re-expression of the same label, and the span-extraction F1 scores in Table 9 cannot serve as independent evidence of a distinct capability. The paper should report the distribution of CM and SI spans per label more carefully, justify why every non-self-harm post is annotated with a casual-mention span, and discuss what the auxiliary tasks add beyond the classification label.
  4. [§7 vs. §8] Section 7 reports that results are "the average of five experimental runs," but Section 8 reports paired t-tests "from 10 runs" with p-values below 0.05; this discrepancy must be resolved because the significance claim for the flagship 0.84 to 0.88 CESM-100 comparison depends on those runs. Beyond the SHC task, the CMSE, SISE, and RG improvements in Table 9 are reported without significance testing, and the statement that significance testing "was not initially performed" for the zero-/few-shot margins does not justify omitting tests for the multitask and ablation comparisons that the paper highlights.
minor comments (7)
  1. [Abstract vs. §3.3] The abstract states that "The SHINES dataset, CESM-100 and codebase are publicly available," but Section 3.3 says "A subset of SHINES annotations is available for review, with the full dataset to be released pending research approval"; these statements should be reconciled before publication.
  2. [§3.3 vs. Ethical Considerations] Section 3.3 attributes the clinical validation to a psychiatrist with "over 12 years of clinical experience," while the Ethical Considerations section says "over 19 years of expertise in mental health"; the two numbers should be aligned.
  3. [§6] The description of the span loss refers to "causal spans" twice; this should read "casual spans" to match the CM terminology used elsewhere.
  4. [Table 9] The caption says the values in parentheses represent variance, but Section 7 says results are averaged over five runs; reporting the number of runs and the run-level F1 distribution would clarify how the paired t-tests in Section 8 relate to the table.
  5. [§A.7 references] The Relevance and Coherence metrics are cited to Teh and Uwasomba (2024), whose title concerns scholarly publication titles and abstracts; please verify that this citation is the correct source for these metrics.
  6. [§8] The model name appears as "Mental-Llama" in the significance-testing paragraph but as "MentaLLaMA" elsewhere; unify the spelling.
  7. [§8.2] The text contains a typo, "for intance," which should be "for instance."

Circularity Check

2 steps flagged · score 7.0 of 10

CESM-100 is built from the same corpus used for evaluation, and rationale-generation metrics reward copying the gold spans that the prompt already supplies, making the central detection and explanation gains partially circular.

  1. fitted input called prediction [Section 4.1 (CESM-100 construction) and Section 7 (train/test split), with the gain reported in Table 9]
    "Section 4.1: "A subset of emojis was extracted from the collected posts (as detailed in last section) related to self-harm and analyzed to understand their usage in these contexts." Section 7: "Our dataset is split 80/20 for training and testing, with synthetic posts used only in training.""

    CESM-100's contextual meanings and CM/SI chance ratings are derived from the same SHINES corpus that is later split into train and test sets to measure CESM-100's contribution. No train/test separation is documented for the emoji extraction in Section 4.1, and Section 5 reports emoji frequencies over the full dataset. Therefore the Table 9 comparison (Llama 3 multitask SHC F1 of 0.84 without CESM-100 vs 0.88 with it) evaluates a resource that has already observed the emoji-label co-occurrences in the test posts. The reported improvement is not a prediction about unseen emoji behavior but a measure of how well CESM-100 encodes corpus-level associations, and the paired t-tests compare two settings that both use the same leaked resource.

  2. self definitional [Table 12 (rationale generation prompt) and Section A.7 (RG metrics)]
    "Table 12: "Using the provided classification and extracted spans, generate a rationale explaining why the post is classified as self-harm or non-self-harm. Consider the spans and the emoji meanings provided." Section A.7: "Checks if all spans are present in the rationale." and "Combines the casual mentions and serious intents into a single text. Uses TF-IDF Vectorizer to transform both the combined text and the rationale into vectors. Computes cosine similarity between these vectors.""

    The rationale-generation prompt supplies the gold CM/SI spans and CESM-100 emoji interpretations as input. The Relevance metric is defined as checking whether all those spans appear in the generated rationale, and Coherence and Semantic Similarity compute vector similarity between the rationale and the combined span text. A model that copies the supplied spans will therefore achieve high RG scores by construction. The reported RG improvements (e.g., SemSim 0.88 for Llama 3) do not demonstrate generalizable explanation quality; they largely measure span-copying of information already present in the prompt. Thus the explanation-quality claim is self-definitional under the chosen metrics.

full rationale

The paper's multitask fine-tuning machinery itself is not circular: the SHC label loss, span extraction losses, and held-out 80/20 split for the classifier are standard. However, two evaluation loops undermine the central empirical claims. First, CESM-100 is constructed from the same posts used for evaluation. Section 4.1 states that emojis were "extracted from the collected posts" and analyzed for self-harm usage, and Section 5's frequency statistics are computed over the full SHINES corpus; the 80/20 split appears only in Section 7, after CESM-100 exists. Consequently, the ablation that credits CESM-100 with the 0.84 to 0.88 F1 gain (Table 9) cannot distinguish a reusable resource from leakage of test-set emoji-label associations. Second, the rationale-generation evaluation is near-tautological: Table 12 injects gold CM/SI spans and CESM-100 interpretations into the prompt, and Section A.7 defines Relevance, Coherence, and Semantic Similarity against exactly those spans, so high RG scores indicate copying rather than independent explanation. The self-citations to Ghosh et al. (2022) and Poria et al. (2021) are methodological and not load-bearing. Separately, the abstract's claim that SHINES is "publicly available" conflicts with Section 3.3's statement that "the full dataset is to be released pending research approval"; this is a correctness and reproducibility concern, not a circularity one. Overall, the core detection and explanation improvements are partly circular as evaluated, warranting a score of 7.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-constructed resources (CESM-100 and span annotations) and domain assumptions about their validity and transferability. No new physical or theoretical entities are introduced. The main free parameter is the CESM-100 Low/Medium/High association levels, which are subjective and not externally benchmarked. The axioms around synthetic-data fidelity, annotation reliability, and automated rationale metrics each carry uncertainty that the paper partially acknowledges.

free parameters (1)
  • CESM-100 emoji intent association levels = Low/Medium/High per emoji (e.g., Coffin: CM Low, SI High; Broken Heart: CM Medium, SI High)
    Hand-assigned by annotators/authors for 100 emojis; these categorical labels are injected into model prompts and drive the reported CESM-100 gain. They are chosen by hand, not learned from held-out data.
assumptions (4)
  • domain assumption Annotator labels and span annotations are reliable ground truth
    Section 3.2 uses majority voting with Fleiss kappa 0.78, but span-level macro-F1 is only 0.66 (CM) and 0.69 (SI), leaving substantial span noise that propagates into training and evaluation.
  • domain assumption Synthetic ChatGPT posts are distributionally similar to real Reddit posts
    Section 3.3 reports annotators can distinguish synthetic from original posts with F1 58%, described as 'above random but not definitive'. Training includes 1000 such posts, which could shift the learned distribution.
  • domain assumption Emoji interpretations in CESM-100 are valid and transferable
    Section 4.1/4.2 builds CESM-100 from emojis extracted from the same collected corpus plus author/annotator expertise, with input from one psychiatrist. No external validation is performed, and the resource is not tested on a disjoint corpus.
  • domain assumption Automated metrics measure rationale/explanation quality
    Section A.7: Relevance requires span containment in the rationale, Coherence and SemSim compare the rationale to the same spans fed into the prompt. No human evaluation of rationales is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Just a Scratch: Enhancing LLM Capabilities for Self-harm Detection through Intent Differentiation and Emoji Interpretation." pith.science (2026). https://pith.science/paper/EKF7RO4I

@misc{pith2026250605073,
  author       = {Pith},
  title        = {Pith review of: Just a Scratch: Enhancing LLM Capabilities for Self-harm Detection through Intent Differentiation and Emoji Interpretation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EKF7RO4I}},
  note         = {Machine review of arXiv:2506.05073}
}
read the original abstract

Self-harm detection on social media is critical for early intervention and mental health support, yet remains challenging due to the subtle, context-dependent nature of such expressions. Identifying self-harm intent aids suicide prevention by enabling timely responses, but current large language models (LLMs) struggle to interpret implicit cues in casual language and emojis. This work enhances LLMs' comprehension of self-harm by distinguishing intent through nuanced language-emoji interplay. We present the Centennial Emoji Sensitivity Matrix (CESM-100), a curated set of 100 emojis with contextual self-harm interpretations and the Self-Harm Identification aNd intent Extraction with Supportive emoji sensitivity (SHINES) dataset, offering detailed annotations for self-harm labels, casual mentions (CMs), and serious intents (SIs). Our unified framework: a) enriches inputs using CESM-100; b) fine-tunes LLMs for multi-task learning: self-harm detection (primary) and CM/SI span detection (auxiliary); c) generates explainable rationales for self-harm predictions. We evaluate the framework on three state-of-the-art LLMs-Llama 3, Mental-Alpaca, and MentalLlama, across zero-shot, few-shot, and fine-tuned scenarios. By coupling intent differentiation with contextual cues, our approach commendably enhances LLM performance in both detection and explanation tasks, effectively addressing the inherent ambiguity in self-harm signals. The SHINES dataset, CESM-100 and codebase are publicly available at: https://www.iitp.ac.in/~ai-nlp-ml/resources.html#SHINES .

Figures

Figures reproduced from arXiv: 2506.05073 by the authors.

Figure 1
Figure 1. LLM Fine-Tuning for Self-Harm Detection and Rationale Generation. The red dotted line encloses the baseline evaluation setup for self-harm detection. • Rationale Generation: The LLM is further prompted to generate rationales that explain the model’s decisions, explicitly referencing CM and SI spans as well as CESM-100-based emoji interpretations. Prompt-Driven Fine-Tuning: The framework utilizes task-specific prompt… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 31 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    AI@Meta. 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card

    2024

  4. [4]

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. Llm2vec: Large language models are secretly powerful text encoders. arXiv preprint arXiv:2404.05961

    arXiv 2024

  5. [5]

    Adrian I Campos, Karin JH Verweij, Dixie J Statham, Pamela AF Madden, Dominique F Maciejewski, Katrina AS Davis, Ann John, Matthew Hotopf, Andrew C Heath, Nicholas G Martin, et al. 2020. Genetic aetiology of self-harm ideation and behaviour. Scientific reports, 10(1):9713

    work page 2020

  6. [6]

    Chi-Keung Chan and Kelly Ka-Wai Chan. 2022. Enhancing the awareness of e-mental health messages: The effects of narrative, emoji, and relevance. In Digital Communication and Learning: Changes and Challenges, pages 87--105. Springer

    work page 2022

  7. [7]

    Neil Cohn, Jan Engelen, and Joost Schilperoord. 2019. The grammar of emoji? constraints on communicative pictorial sequencing. Cognitive research: principles and implications, 4:1--18

    work page 2019

  8. [8]

    Marcel Danesi. 2022. Emotional wellbeing and the semiotic translation of emojis. In Exploring the Translatability of Emotions: Cross-Cultural and Transdisciplinary Encounters, pages 323--344. Springer

    work page 2022

Show all 39 references
  1. [9]

    Jason Davies, Mark McKenna, Kate Denner, Jon Bayley, and Matthew Morgan. 2024. The emoji current mood and experience scale: the development and initial validation of an ultra-brief, literacy independent measure of psychological health. Journal of Mental Health, 33(2):218--226

    2024

  2. [10]

    Saahil Deshpande and Jim Warren. 2021. Self-harm detection for mental health chatbots. In Public Health and Informatics, pages 48--52. IOS Press

    2021

  3. [11]

    Thomas G Dietterich. 1998. Approximate statistical tests for comparing supervised classification learning algorithms. Neural computation, 10(7):1895--1923

    1998

  4. [12]

    Xiruo Ding, Kevin Lybarger, Justin Tauscher, and Trevor Cohen. 2022. Improving classification of infrequent cognitive distortions: domain-specific model vs. data augmentation. In Proceedings of the 2022 conference of the North American chapter of the association for computatio...

    2022

  5. [13]

    David Duki \'c and Jan S najder. 2024. Looking right is sometimes right: Investigating the capabilities of decoder-only llms for sequence labeling. In Findings of the Association for Computational Linguistics ACL 2024, pages 14168--14181

    2024

  6. [14]

    Jon Ennis, Rosemary A Barnes, Sidney Kennedy, and Dvora D Trachtenberg. 1989. Depression in self-harm patients. The British journal of psychiatry, 154(1):41--47

    1989

  7. [15]

    Manuel Faysse, Gautier Viaud, C \'e line Hudelot, and Pierre Colombo. 2023. Revisiting instruction fine-tuned model evaluation to guide industrial applications. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9033--9048

    2023

  8. [16]

    Rudolf Flesch. 2007. Flesch-kincaid readability test. Retrieved October, 26(3):2007

    2007

  9. [17]

    Soumitra Ghosh, Asif Ekbal, and Pushpak Bhattacharyya. 2022. A multitask framework to detect depression, sentiment and multi-label emotion from suicide notes. Cognitive Computation, 14(1):110--129

    2022

  10. [18]

    Declan Grabb, Max Lamparth, and Nina Vasan. 2024. Risks from language models for automated mental healthcare: Ethics and structure for implementation. In Proceedings of the 2024 AAAI/ACM Conference on AI, Ethics, and Society, pages 519--519

    2024

  11. [19]

    Vandita Grover. 2022. Exploiting emojis in sentiment analysis: A survey. Journal of The Institution of Engineers (India): Series B, 103(1):259--272

    2022

  12. [20]

    Lin Gui, Ruifeng Xu, Dongyin Wu, Qin Lu, and Yu Zhou. 2018. Event-driven emotion cause extraction with corpus construction. In Social Media Content Analysis: Natural Language Processing and Beyond, pages 145--160. World Scientific

    2018

  13. [21]

    Colin ME Halverson, Claire E Donnelly, Michael Weiner, and Joy L Lee. 2023. Content analysis of emoji and emoticon use in clinical texting systems. JAMA Network Open, 6(6):e2318140--e2318140

    2023

  14. [22]

    Thomas F Heston. 2023. Safety of large language models in addressing depression. Cureus, 15(12)

    2023

  15. [23]

    Wassily Hoeffding. 1994. Probability inequalities for sums of bounded random variables. The collected works of Wassily Hoeffding, pages 409--426

    1994

  16. [24]

    E David Klonsky, Alexis M May, and Catherine R Glenn. 2013. The relationship between nonsuicidal self-injury and attempted suicide: converging evidence from four samples. Journal of abnormal psychology, 122(1):231

    2013

  17. [25]

    Kai Xiang Lim, Olakunle Ayokunmi Oginni, Kaili Rimfeld, Jean-Baptiste Pingault, and Fr \"u hling Rijsdijk. 2022. Investigating the causal risk factors for self-harm by integrating mendelian randomisation within twin modelling. Behavior genetics, 52(6):324--337

    2022

  18. [26]

    Mounica Maddela, Megan Ung, Jing Xu, Andrea Madotto, Heather Foran, and Y-Lan Boureau. 2023. Training models to generate, recognize, and reframe unhelpful thoughts. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Paper...

    2023

  19. [27]

    Ryan K McBain, Jonathan H Cantor, Li Ang Zhang, Olesya Baker, Fang Zhang, Alyssa Halbisen, Aaron Kofner, Joshua Breslau, Bradley Stein, Ateev Mehrotra, et al. 2025. Competency of large language models in evaluating appropriate responses to suicidal ideation: Comparative study....

    2025

  20. [28]

    Soujanya Poria, Navonil Majumder, Devamanyu Hazarika, Deepanway Ghosal, Rishabh Bhardwaj, Samson Yu Bai Jian, Pengfei Hong, Romila Ghosh, Abhinaba Roy, Niyati Chhaya, et al. 2021. Recognizing emotion cause in conversations. Cognitive Computation, 13:1317--1332

    2021

  21. [29]

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383--2392

    2016

  22. [30]

    Abigail Emma Russell, Gibran Hemani, Hannah J Jones, Tamsin Ford, David Gunnell, Jon Heron, Carol Joinson, Paul Moran, Caroline Relton, Matthew Suderman, et al. 2021. An exploration of the genetic epidemiology of non-suicidal self-harm and suicide attempt. BMC psychiatry, 21(1):207

    2021

  23. [31]

    Testing the limits of unified sequence to sequence llm pretraining on diverse table data tasks

    Soumajyoti Sarkar and Leonard Lausen. Testing the limits of unified sequence to sequence llm pretraining on diverse table data tasks. In NeurIPS 2023 Second Table Representation Learning Workshop

    2023

  24. [32]

    Robert L Spitzer, Jacob Cohen, Joseph L Fleiss, and Jean Endicott. 1967. Quantification of agreement in psychiatric diagnosis: A new approach. Archives of General Psychiatry, 17(1):83--87

    1967

  25. [33]

    Phoey Lee Teh and Chukwudi Festus Uwasomba. 2024. Impact of large language models on scholarly publication titles and abstracts: A comparative analysis. Journal of Social Computing, 5(2):105--121

    2024

  26. [34]

    Anuradha Welivita and Pearl Pu. 2023. Boosting distress support dialogue responses with motivational interviewing strategy. In Findings of the Association for Computational Linguistics: ACL 2023, pages 5411--5432

    2023

  27. [35]

    Janis Whitlock, Jennifer Muehlenkamp, John Eckenrode, Amanda Purington, Gina Baral Abrams, Paul Barreira, and Victoria Kress. 2013. Nonsuicidal self-injury as a gateway to suicide in young adults. Journal of adolescent health, 52(4):486--492

    2013

  28. [36]

    Xuhai Xu, Bingsheng Yao, Yuanzhe Dong, Saadia Gabriel, Hong Yu, James Hendler, Marzyeh Ghassemi, Anind K Dey, and Dakuo Wang. 2024. Mental-llm: Leveraging large language models for mental health prediction via online text data. Proceedings of the ACM on Interactive, Mobile, We...

    2024

  29. [37]

    Kailai Yang, Tianlin Zhang, Ziyan Kuang, Qianqian Xie, Jimin Huang, and Sophia Ananiadou. 2024 a . Mentallama: interpretable mental health analysis on social media with large language models. In Proceedings of the ACM on Web Conference 2024, pages 4489--4500

    2024

  30. [38]

    Zi Yun Yang, Ziqing Zhang, and Yisong Miao. 2024 b . The elco dataset: Bridging emoji and lexical composition. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 15899--15909

    2024

  31. [39]

    H Yu and Stephen McGuinness. 2024. An experimental study of integrating fine-tuned llms and prompts for enhancing mental health support chatbot system. Journal of Medical Artificial Intelligence, pages 1--16

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.