CoChat

Can Large Language Models Transform Computational Social Science?

Open access
Ziems, Caleb, Held, William, Shaikh, Omar, Chen, Jiaao, Zhang, Zhehao, Yang, Diyi · 2024 · Computational Linguistics

The DOI matches its Crossref record; 320 of the 324 references it cites were found in a registry or an index; none of the 111 sources its claims cite is checked for support yet.

Abstract

Large language models (LLMs) are capable of successfully performing many language processing tasks zero-shot (without training data). If zero-shot LLMs can also reliably classify and explain social phenomena like persuasiveness and political ideology, then LLMs could augment the computational social science (CSS) pipeline in important ways. This work provides a road map for using LLMs as CSS tools. Towards this end, we contribute a set of prompting best practices and an extensive evaluation pipeline to measure the zero-shot performance of 13 language models on 25 representative English CSS benchmarks. On taxonomic labeling tasks (classification), LLMs fail to outperform the best fine-tuned models but still achieve fair levels of agreement with humans. On free-form coding tasks (generation), LLMs produce explanations that often exceed the quality of crowdworkers’ gold references. We conclude that the performance of today’s LLMs can augment the CSS research pipeline in two ways: (1) serving as zero-shot data annotators on human annotation teams, and (2) bootstrapping challenging creative generation tasks (e.g., explaining the underlying attributes of a text). In summary, LLMs are posed to meaningfully participate in social science analysis in partnership with humans.

Overview

CoChat Summary

TL;DR

This paper evaluates whether zero-shot large language models can augment the Computational Social Science pipeline by benchmarking 13 LLMs on 25 English CSS tasks. It concludes that LLMs can assist with annotation and generation but cannot fully replace fine-tuned models or human analysis.

Key findings

  • On classification, zero-shot LLMs fail to outperform the best fine-tuned models but still achieve fair agreement with humans: 8/17 tasks reach moderate-to-good agreement (κ = 0.40–0.65), while 5/17 (29.5%) show poor agreement. Best zero-shot performance reaches 77.4 F1 on misinformation detection, with κ = 0.55 versus the original paper’s κ = 0.51 inter-human agreement.
  • On generation, LLMs often exceed crowdworkers’ gold references: model generations outrank the dataset’s gold human reference at least 38% of the time across five tasks. For COVIDET, Misinformation Explanation, and Social Bias Inference, gpt-3.5-turbo produces more faithful, coherent, and fluent text than both gold references and the fine-tuned baseline; for Positive Reframing and Figurative Language explanation, text-davinci-003 matches gold-standard levels.
  • Model selection matters: larger instruction-tuned open-source FLAN models are preferable for classification, while gpt-3.5-turbo is the clear leading LLM for generation. FLAN performance scales nearly logarithmically, with a median improvement of 5.0 absolute percentage points per order-of-magnitude size increase; OpenAI models do not monotonically benefit from scale, and RLHF improves average F1 by 3.5 absolute points over text-davinci-002.
  • Performance is not tied to academic discipline, but document-level tasks are hardest, with all F1 scores near or below 50%; complex expert taxonomies also perform poorly, e.g., implicit hate 32.3 F1, empathy 35.9 F1, and dialect feature detection 32.9 F1.

Methods

  • The authors evaluate zero-shot performance of 13 language models on 25 representative English CSS benchmarks: 10 utterance-level classification tasks, 6 conversation-level tasks, 4 document-level tasks, and 5 generation tasks. Models include FLAN-T5 sizes from 80M to 11B, FLAN-UL2 (20B), OpenAI GPT-3 text-001/002/003, gpt-3.5-turbo, and GPT-4.
  • They use prompt engineering guidelines, generate five semantically equivalent prompt perturbations per task, set temperature to zero, and constrain decoding to valid output classes or use logit bias. Classification and structured parsing are evaluated with macro F1 averaged over prompt perturbations; supervised baselines use RoBERTa-large for classification and T5-base for generation, tuned by grid search and averaged over three random seeds.
  • For each task, they evaluate a class-stratified sample of at most 500 instances from the designated test set. Generation is evaluated by domain experts on 1–5 Likert scales for faithfulness, relevance, coherence, and fluency, and by ranking model outputs against gold human references; ranking annotations were performed by two authors blind, with unanimous voting.

Limitations

  • The paper states that its task selection is not exhaustive; sensitive domains such as mental health and cultural studies are largely outside scope, and available data resources are dominated by standard dialects and WEIRD populations. It also notes data leakage concerns for proprietary models, that it explores predictive rather than causal explanations, that bias and fairness could not be systematically analyzed with available datasets, and that proprietary LLMs carry opacity, privacy, and intellectual-property risks.
  • The paper states that generation diversity may be limited and that human evaluation is costly and hard to reproduce; it also notes that exact model sizes and training data for OpenAI models are trade secrets, so scaling analyses rely on estimates.
  • Evident from the text: classification results are based on class-stratified samples capped at 500 instances per task, prompt variance is addressed with only five perturbations, and the human ranking evaluation used the authors as blind annotators rather than independent crowdworkers. Some dataset categories with fewer than 10 examples were omitted from evaluation, and automatic generation metrics such as BLEU, BERTScore, and BLEURT produced uninterpretable scores.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1162/coli_a_00502

Cites

320 of 324 references verified0 citation errors · 4 not checkable

Statistics

Nothing to recompute.

Figures

Image 1
CC-BY-NC-ND · From the PDF
Image 2
CC-BY-NC-ND · From the PDF
Image 3
CC-BY-NC-ND · From the PDF
Image 4
CC-BY-NC-ND · From the PDF
Figure 5 Human-AI Collaboration can improve the efficiency and reliability of text analysis. In this misinformation example, the LLM helps scale up annotation while reducing variance in the gold labels. Human annotation serves as validation for model-provided annotations.
Figure 5 Human-AI Collaboration can improve the efficiency and reliability of text analysis. In this misinformation example, the LLM helps scale up annotation while reducing variance in the gold labels. Human annotation serves as validation for model-provided annotations.CC-BY-NC-ND · From the PDF

Verification

Claims the paper makes

0 Valid · 14 Not valid · 77 Not checked · 179 Own results and interpretations

114 more claims

Create a free account to see all 270 claims and to verify the 92 references still open.

See all claims — free account
  • It may state two separate claims as one. 80
  • Probably not a substantial claim. 91
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

1 more claim about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

4 of 4

Citations

Cites

Show all 324 references

Research works printed but not linked: 12

  1. Cialdini, 2003, Influence
  2. 73. Dang, Hoa Trang. 2005. Overview of duc 2005. In Proceedings of the document understanding conference, volume 2005, pages 1-12, Citeseer.
  3. 77. Ziems, Held, Shaikh, Chen, Zhang, Yang Can LLMs Transform CSS? Workshop held in conjunction with ACL 2016 (LAW-X 2016), pages 188-198, Association for Computational Linguistics, Berlin, Germany.
  4. 100. Garfinkel, Harold. 2016. Studies in ethnomethodology. In Social Theory Re-Wired. Routledge, pages 85-95.
  5. 109. Gross, Justin H, Brice Acree, Yanchuan Sim, and Noah A Smith. 2013. Testing the etch-a-sketch hypothesis: a computational analysis of mitt romney's ideological makeover during the 2012 primary vs. general elections. In APSA 2013 Annual Meeting Paper, American Political Science Association 2013 Annual Meeting.
  6. 152. Kuhn, Thomas S. 1962. The structure of Scientific Revolutions. The University of Chicago Press.
  7. 156. Ziems, Held, Shaikh, Chen, Zhang, Yang Can LLMs Transform CSS? Conference on Human Factors in Computing Systems, CHI 2015, Seoul, Republic of Korea, April 18-23, 2015, pages 3739-3748, ACM.
  8. 172. Luo, Yiwei, Dallas Card, and Dan Jurafsky. 2020b. Detecting stance in media on global warming. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3296-3315, Association for Computational Linguistics, Online. · DOI
  9. 188. Murray, Michael et al. 2015. Narrative psychology. Qualitative psychology: A practical guide to research methods, pages 85-107.
  10. Omitaomu, 2022, Empathic conversations: A multi-level dataset of contextualized conversations, ArXiv preprint
  11. Rothbaum, 2000, Cognitive-behavioral therapy, 320
  12. 227. Sacks, Harvey. 1992. Lectures on conversation: Volume i. Malden, Massachusetts: Blackwell.

Cited by

Items 101–200 of 609

Metadata

Authors
Ziems, C., Held, W., Shaikh, O., Chen, J., Zhang, Z., & Yang, D.
Affiliations
  • Stanford University (US)
  • Georgia Institute of Technology (US)
  • Dartmouth College (US)
Venue
Computational Linguistics
Publisher
MIT Press
Published
2024
Volume
50
Issue
1
Pages
237-291
ISSN
  • 0891-2017 (ISSN-L)
  • 1530-9312
OpenAlex
W4389636360
Cited by
585
References
311
Citations by year
2026: 201 · 2025: 197 · 2024: 146
FWCI
271.63
Type
journal article
Language
en
Open access
Locations
  • Computational Linguistics (publishedVersion, cc-by-nc-nd)
  • arXiv (Cornell University) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University)
Funders
National Science Foundation (2247357, 2308994, 2039655)
Topics
  • Social Sciences › General Social Sciences › Computational and Text Analysis Methods
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Sentiment Analysis and Opinion Mining
Keywords
large language models, computational social science, zero-shot learning, persuasion detection, text classification, text generation, manual annotation, evaluation pipeline

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-21, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-21 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.