diff --git "a/20240921/2409.09467v2.json" "b/20240921/2409.09467v2.json" new file mode 100644--- /dev/null +++ "b/20240921/2409.09467v2.json" @@ -0,0 +1,662 @@ +{ + "title": "Keeping Humans in the Loop: Human-Centered Automated Annotation with Generative AI", + "abstract": "Automated text annotation is a compelling use case for generative large language models (LLMs) in social media research.\nRecent work suggests that LLMs can achieve strong performance on annotation tasks; however, these studies evaluate LLMs on a small number of tasks and likely suffer from contamination due to a reliance on public benchmark datasets.\nHere, we test a human-centered framework for responsibly evaluating artificial intelligence tools used in automated annotation. We use GPT-4 to replicate 27 annotation tasks across 11 password-protected datasets from recently published computational social science articles in high-impact journals. For each task, we compare GPT-4 annotations against human-annotated ground-truth labels and against annotations from separate supervised classification models fine-tuned on human-generated labels. Although the quality of LLM labels is generally high, we find significant variation in LLM performance across tasks, even within datasets. Our findings underscore the importance of a human-centered workflow and careful evaluation standards: Automated annotations significantly diverge from human judgment in numerous scenarios, despite various optimization strategies such as prompt tuning. Grounding automated annotation in validation labels generated by humans is essential for responsible evaluation.", + "sections": [ + { + "section_id": "1", + "parent_section_id": null, + "section_name": "Introduction", + "text": "###figure_1### ###figure_2### Scholars studying social media routinely employ natural language processing (NLP) tools to analyze large quantities of text data. Classification is among the most common NLP tasks, allowing researchers to describe the text of social media posts at scale. Social media researchers have begun to deploy generative LLMs for annotation for important tasks, such as identifying hate speech (Kumarage, Bhattacharjee, and Garland 2024 ###reference_b30###), public attitudes towards vaccines (Espinosa and Salath\u00e9 2024 ###reference_b13###), and measuring news outlet credibility (Yang and Menczer 2023 ###reference_b54###). Classification tasks depend on high-quality, manually-labeled text data for training and validation of supervised and semi-supervised language models (Grimmer, Roberts, and Stewart 2022 ###reference_b18###). Computational social scientists typically use the labeled text data from these models for descriptive purposes or to serve as control, predictor, or outcome variables in statistical analyses (Egami et al. 2022 ###reference_b12###; Knox, Lucas, and Cho 2022 ###reference_b28###; Roberts, Stewart, and Nielsen 2020 ###reference_b45###).\nHere, we explore the potential of generative large language models (LLMs) to automate the manual annotation process as a few-shot classification problem (Brown et al. 2020 ###reference_b4###). LLMs are faster and cheaper than human annotators and do not suffer from human shortcomings such as limited attention span or fatigue (Grimmer and Stewart 2013 ###reference_b19###; Neuendorf 2016 ###reference_b39###). As a result, LLMs may be valuable tools for annotation.\nIn Figure 1, we propose a human-centered workflow for automated annotation. Despite inevitable imperfections of human annotators, we argue that deployment of artificial intelligence (AI) technology for automated annotation must remain grounded in human judgment. Human-centered artificial intelligence (HCAI) is a central principle of responsible AI and a critical component to developing automated technology that is Reliable, Safe, and Trustworthy (RST) (Falco et al. 2021 ###reference_b14###; Shneiderman 2020 ###reference_b47###). Without a human-centered approach that relies on human validation data, researchers cannot identify and measure biases in annotation performance.\nA growing body of research on automated annotation claims that generative LLMs can match or exceed human performance on annotation tasks (Chiang and Lee 2023 ###reference_b7###; Ding et al. 2022 ###reference_b11###; He et al. 2023 ###reference_b20###; Mellon et al. 2022 ###reference_b36###; Pan et al. 2023 ###reference_b41###; Thapa, Naseem, and Nasim 2023 ###reference_b49###; T\u00f6rnberg 2023a ###reference_b50###, b ###reference_b51###; Zhu et al. 2023 ###reference_b56###). In tension with a human-centered approach, some of this research is framed to evaluate whether LLMs are \u201cbetter\u201d than human annotators (e.g. Huang, Kwak, and An 2023 ###reference_b22###). For instance, Gilardi, Alizadeh, and Kubli (2023 ###reference_b17###) conclude that LLMs outperform typical human annotators: \u201c[t]he evidence is consistent across different types of texts and time periods. It strongly suggests that ChatGPT may already be a superior approach compared to crowd annotations on platforms such as MTurk.\u201d Others argue that automated annotations by LLMs may be higher quality than annotations by human experts (T\u00f6rnberg 2023a ###reference_b50###) and, as a result, \u201chuman coding hence cannot be treated as an unquestioned gold standard\u201d (T\u00f6rnberg 2023b ###reference_b51###). The implication is that generative LLMs could replace human annotators.\nDespite these claims, it is unclear whether LLM performance in previous studies can generalize to other datasets and tasks. Most studies in this vein analyze a few insular tasks on a small number of datasets test LLM performance on publicly available benchmark datasets (Wang et al. 2021 ###reference_b52###; Ziems et al. 2023 ###reference_b57###), which are plausibly affected by data leakage and contamination.111Contamination means that the datasets may be included in the LLM\u2019s pretraining data. As a result, strong performance may reflect memorization, which will not generalize to datasets and tasks that are not included in the LLM\u2019s training data. Aside from concerns about contamination, errors in automated annotation are likely to be correlated across samples, which may lead to systematic biases in labels (Ashwin, Chhabra, and Rao 2023 ###reference_b3###). Others have noted that inconsistency in ChatGPT\u2019s annotations means they are unreliable (Ollion et al. 2023 ###reference_b40###; Reiss 2023 ###reference_b44###). These concerns suggest that rigorous, task-by-task validation is essential for practitioners relying on LLMs for automated annotation.\nIf the strong performance demonstrated in recent work generalizes to novel text annotation tasks in social science, then validation may not be a significant concern. If performance does not generalize to a wide range of tasks, then researchers risk generating inaccurate labels that will introduce bias in downstream description or analysis. Researchers who use LLMs for automated annotation without human-centered validation assume LLM annotations align with human judgment about the conceptual meaning of categories. Without ground-truth labels, these assumptions cannot be tested. We argue that using LLMs for automated annotation responsibly and effectively requires a human-centered approach.\nIn this paper, we highlight the importance of grounding automated annotation in human judgment by testing the generalized performance of automated annotation on tasks and datasets less likely to be affected by contamination and leakage. We use generative LLMs to replicate 27 different manual annotation procedures from 11 non-public datasets from computational social science (CSS) articles recently featured in high-impact publications. The replicated tasks represent a range of real-world applications in CSS. Each of the original datasets includes human-labeled annotations that we treat as ground-truth.222We consider the human labels from the original studies as ground truth because the annotation procedures underwent independent review and scrutiny during the peer-review process and were determined to be quality enough for publication in a high-impact journal. All of the annotation procedures in the original studies employed either multiple annotators or an expert coder. We discuss the human annotation procedures from the original studies in Appendix B. Rather than testing whether LLMs \u201coutperform\u201d human annotators, we explore how well they can approximate human judgment. We ask the following question: Can generative AI reasonably approximate human annotations in computational social science?\nFor each task, we provide GPT-4333We selected GPT-4 because it was the highest performing generative LLM at the time of our analyses. with detailed instructions to label text samples into conceptual categories outlined in the original study. We then conduct direct label-to-label comparisons between the GPT-4 annotations and the human annotations in the original study. To compare these results to a baseline, we also fine-tune several supervised text classifiers for every task. Unlike the GPT-4 few-shot approach, these supervised models were fine-tuned with human-generated labels only. Due to its low cost, ease of application, and high performance, we fine-tune BERT classifiers (Devlin et al. 2019 ###reference_b10###) on varying training sample sizes and use them to benchmark GPT-4\u2019s performance. After our primary investigation, we conduct a series of secondary analyses aimed at exploring ways to improve automated annotation performance. These experiments include testing prompt optimization, temperature tuning, and a novel strategy to identify lower confidence automated annotations.\nWe identify four main contributions from our results:\nGPT-4 text annotation performance is promising but inconsistent across tasks and datasets. Despite excelling on many tasks, automated annotation significantly diverges from human judgment in numerous scenarios. Across all replicated tasks, we find a median accuracy of 0.850 and a median F1 of 0.707. Nevertheless, nine of the 27 tasks were below 0.5 on precision or recall. For a full one third of the sampled tasks, automated annotation would result in labels for which false positions outnumber true positives or more than half the true positives are missed. Grounding automated annotation in validation labels generated by humans is essential for responsible evaluation.\nGPT-4 performance is significantly stronger in recall than precision. For 20 of the 27 tasks, recall exceeds precision. Thus, one valuable use case of automated annotation may be as a first stage in a multi-stage annotation pipeline, where human annotators manually review all positive class instances from the LLM outputs.\nVarious strategies to improve automated annotation performance, including manual prompt optimization and tuning LLM hyperparameters like temperature, improve annotation performance only marginally. These strategies are unlikely to solve the inconsistencies in poor LLM annotation performance at this time.\nWith adequate training data, supervised encoder-only LLM classifiers surpass GPT-4 performance. GPT-4 only outperforms supervised classifiers when there are minimal training samples." + }, + { + "section_id": "2", + "parent_section_id": null, + "section_name": "Data and Replication Procedures", + "text": "" + }, + { + "section_id": "2.1", + "parent_section_id": "2", + "section_name": "Data", + "text": "The data used in our replication analyses include 27 manual annotation tasks from 11 CSS articles recently published in high-impact journals.444Table A1 lists the complete list of the datasets and Table A2 details each annotation task. These tasks were selected to represent a range of annotation tasks in contemporary CSS research, drawing from research appearing in interdisciplinary publications (e.g., Science Advances and Proceedings of the National Academy of Sciences) as well as high-impact field journals in political science (e.g., American Political Science Review and American Journal of Political Science) and psychology (e.g., Journal of Personality and Social Psychology). We identified these articles by searching journals for articles that utilized manual annotation procedures in their research design. All replicated annotation procedures appeared in articles that were published within the last three years and included replication data in a password-protected data repository.\nWe treat each label in the original datasets as a separate binary classification task. While the majority of the classification tasks we replicate were originally binary classification tasks, we decompose the remaining multiclass classification tasks into binary tasks. Decomposing classification problems into a series of binary classification problems is a simple, common approach for multiclass classification (Allwein, Schapire, and Singer 2000 ###reference_b1###; Joshi et al. 2017 ###reference_b25###; Lorena, De Carvalho, and Gama 2008 ###reference_b33###). This approach has several advantages for our analysis. First, it harmonizes how we execute and evaluate each replicated task, allowing for a standardized classification workflow across tasks. Second, treating each category as a separate binary classification task allows for more granular error analysis for LLM performance. Given our focus on evaluating LLMs at text annotation tasks, decomposing single complex tasks into several simpler tasks gives more opportunities for insight into heterogeneity in LLM annotation performance.\nOur tasks cover varying degrees of class imbalance: Across all tasks, the minimum positive class frequency is 0.04%, the median is 24.5%, and the maximum is 70.7%. The sources of the human labels are representative of common approaches to human annotation: 22.2% of tasks were annotated by crowdsourced workers, 37% by experts, and 40.8% by research assistants. The annotation tasks reflect an extensive range of CSS applications, from identifying whether Cold War-era texts pertained to political affairs or military matters (Schub 2022 ###reference_b46###) to analyzing open-ended survey responses to classify how people conceptualize where beliefs come from (Cusimano and Goodwin 2020 ###reference_b8###).\nBecause we rely exclusively on datasets stored in password-protected data archives (e.g., Dataverse) or datasets secured through direct outreach to authors, our replication analyses are unlikely to be affected by contamination and data leakage. Task and test data contamination are significant, yet understudied, challenges inherent to evaluating LLMs (Brown et al. 2020 ###reference_b4###; Li and Flanigan 2023 ###reference_b31###; Magar and Schwartz 2022 ###reference_b34###). Because LLMs are trained on huge volumes of text data, using digital resources such as the Common Crawl, they are able to \u201cmemorize\u201d publicly available information during the training process. This issue is compounded due to the fact that the highest performing models (e.g., GPT-4) are closed-source LLMs and the proprietors do not release details on pre-training data. Although open-source models provide more detail on their training data, extracting data from crawled websites can be difficult and prone to drift over time (Li and Flanigan 2023 ###reference_b31###). By analyzing public data and tasks released before and after the LLM pre-training period, Li and Flanigan (2023 ###reference_b31###) determine that contamination is a widespread evaluation issue and conclude that,\u201c[d]ue to task contamination, closed-sourced models may demonstrate inflated performance in zero-shot or few-shot evaluation, and are therefore not trustworthy baselines in these settings.\u201d\nPrior research that examines automated annotation on a range of tasks mainly evaluates their models on popular, publicly available benchmarks released before the LLM training process (Wang et al. 2021 ###reference_b52###; Ziems et al. 2023 ###reference_b57###). Therefore, it is not clear whether these findings will generalize to non-contaminated tasks, which are more representative of CSS tasks." + }, + { + "section_id": "2.2", + "parent_section_id": "2", + "section_name": "Replication Procedures", + "text": "We follow the four-step human-in-the-loop workflow described in Figure 1 to replicate each annotation task. This workflow centers human judgment by using human ground-truth labels to guide prompt optimization and to evaluate label quality.\nIn step 1, we create instructions (i.e., a qualitative codebook) for each task that delineates the conceptual categories of interest.555Because LLMs respond to natural language prompts, a clear set of instructions is essential for LLM annotation tasks. A rich literature spanning both qualitative and quantitative social science offers guidance on how to develop a codebook and classify text data based on relevant concepts (e.g., Krippendorff 2018 ###reference_b29###).\nWe develop the LLM prompts from the exact human annotator instructions used in the original research design. If the instructions were not directly available, we either quote or paraphrase text from the article or supplementary materials that describes the concepts of interest.666We do not observe any relationship between LLM performance and whether or not the verbatim wording used to generate labels was available. In the supplementary material, we provide all prompt instructions.\nFor step 2, we have the LLM label a randomly selected subset (n=250) of text samples already labeled by humans using the human codebook as a prompt.777If humans and LLMs label the data using different codebooks, there could be a conceptual gap between the two annotation instructions, which could lead to a threat in content validity (Jacobs and Wallach 2021 ###reference_b23###). To ground automated annotation in human judgment, we then calculate classification performance metrics (i.e., accuracy, recall, precision, and F1) by comparing the generative LLM predicted labels against the ground-truth human labels.\nIn step 3, we refine the prompts by reviewing patterns in errors made by the LLM and by incorporating examples from incorrect LLM classifications.888One concern at this stage is overfitting the prompt instructions to this particular subset of data. Moreover, if fundamental changes were made to the prompt instructions, it may be prudent to have the humans re-label the original text samples as well. To do so, we use the same 250 samples from step 2. This step is intended to optimize the quality of the prompt instructions by making LLM judgment align as closely as possible to human judgment.\nIn step 4, we use the generative LLM to label a separate set of 1,000 held-out text samples with the updated prompts. We assess final performance against these held-out samples. Performance on this held-out set can be used to determine whether an LLM can be reasonably utilized for automated annotation. The code made available in our GitHub repository offers a simple and efficient way to implement the procedures outlined in the above workflow.999See: https://github.com/npangakis/gpt\u02d9annotate\nAcross all 27 tasks, we annotate slightly over 75,000 text samples using OpenAI\u2019s GPT-4 API.101010Scholars have raised concerns about relying on LLMs in social science research due to the constant evolution of LLM software, the black-box nature of how LLMs process queries, and ambiguity about the training data that underlies LLMs. Open-source LLMs represent a viable solution to many of these problems (Spirling 2023 ###reference_b48###). The workflow outlined here is LLM-agnostic and could be adapted to any open-source LLM. The overall cost was under $500 USD. On average, a dataset with 1,000 text samples took approximately 1 hour to label using GPT-4. Together, the low cost and relatively rapid speed demonstrate the potential value of LLM-augmented annotation in CSS.\nTo set baselines for model performance, we also replicate each of the annotation tasks as a supervised text classification problem using an encoder-only LLM. For our analyses, we classify the same held-out set from step 4 above using BERT models fine-tuned on varying sample sizes of human-labeled text samples. To tune each model\u2019s hyperparameters, we train 18 separate classifiers with varying hyperparameters and select the model with the highest F1 performance.111111We optimize the learning rate, the batch size, and the number of epochs. We elaborate on this process in Appendix C. In total, we trained over 500 supervised LLM classifiers to better understand the performance of GPT-4 automated annotation relative to alternative classification strategies.\nFinally, after conducting our main analyses, we test a variety of approaches to improve automated annotation performance, including prompt optimization, temperature tuning, and a novel selection strategy. We also explore whether GPT-4 performance is change over time. These secondary experiments aim to understand whether there may be specific ways to overcome poor performing tasks." + }, + { + "section_id": "3", + "parent_section_id": null, + "section_name": "Results", + "text": "Automated annotation performance metrics are shown in Table 1 ###reference_###. The performance metrics reported are based on 1,000 \u201cheld out\u201d text samples per task (i.e., text samples not used in the prompt optimization process from step 2 of the workflow). Across the 27 tasks, GPT-4 achieved a median F1 score of 0.707 and median accuracy of 0.85.121212Prior work (e.g., Gilardi, Alizadeh, and Kubli 2023) also reports LLM \u201cintercoder agreement,\u201d which indicates how consistently an LLM labels a text sample. Yet, this measure cannot distinguish between inconsistency due to poor performance, substantive ambiguity, or arbitrary randomness governed by a \u201ctemperature\u201d parameter. As a result, we do not leverage this metric for automated annotation evaluation. Instead, we exploit this measure as a gauge of annotation \u201cconfidence,\u201d which we explore in our \u201cExploiting Uncertainty in LLM Labels\u201d section. Thus, automated annotation is a viable strategy for a significant portion of the replicated tasks.131313Appendix D investigates an open-source alternative to the methods explored here.\n###figure_3### Figure 2 ###reference_### shows precision and recall for each annotation task. On eight of the 27 tasks, the LLM achieves remarkably strong performance with precision and recall both exceeding 0.7. Nevertheless, automated annotations diverged from human judgment in a non-trivial number of cases. Specifically, nine of the 27 tasks had either precision or recall below 0.5\u2014and three tasks had both precision and recall below 0.5. Put otherwise, for a full one-third of tasks, the LLM either missed at least half of the true positive cases, had more false positives than true positives, or both. At worst, the aggregate performance ranged as low as an F1 score of 0.06. These results are not driven by one data set or type of annotation source. Moreover, LLM performance varied substantially across tasks within a single dataset. In the most extreme example, F1 ranged from 0.259 to 0.811 on two separate tasks a single dataset (political speech data from Card et al. 2022 ###reference_b5###), a difference of 0.552. These results demonstrate the variability of automated annotation performance and, accordingly, underscore the need for task-specific human validation.\n###figure_4### Although performance was strong across many tasks, especially those akin to summarization (for instance M\u00fcller 2022 ###reference_b38###), it suffered on more abstract tasks. For instance, poor performance correlated with tasks relying on nuanced cultural knowledge, such as differentiating between national, racial, and religious identities (Hopkins, Lelkes, and Wolken 2024 ###reference_b21###). Subpar performance also occurred when the task required inferences about conversational context (Peng, Romero, and Horvat 2022 ###reference_b43###). Although we cannot be certain why certain tasks performed poorly, it is possible that, in general, performance was lower than prior studies because the data sets analyzed here were stored in password-protected archives and thus not affected by memorization. In Appendix E, we conduct additional analyses investigating how performance varies across different conceptualizations of task and text difficulty.\nAs is clear from Figure 2 ###reference_###, LLM performance is stronger on recall than precision for 20 of the 27 tasks. Overall, the median recall (0.83) is significantly higher than the median precision (0.65). Accordingly, automated annotation may be especially valuable as a first stage in multi-stage annotation pipelines, providing high-recall preliminary labels that are refined in subsequent stages by models optimized for precision or by additional manual review. Following a human-machine collaborative framework (Li et al. 2023 ###reference_b32###), one future application may be to use generative AI as a technique to create more balanced classes before the manual annotation process. This would involve an LLM labeling a large number of text samples and then downsampling to create class balance for the initial human annotation stage." + }, + { + "section_id": "3.1", + "parent_section_id": "3", + "section_name": "Manual Prompt Optimization", + "text": "For each of the 27 annotation tasks, we made iterative human-in-the-loop updates to the instructions to optimize the prompt for accurate annotations. Figure 3 shows the distributions of change in performance metrics after updating the prompt and re-annotating the same text samples. This analysis tests whether and how the prompt update process affects LLM annotation, holding constant the data and conceptual categories.\nIn most cases, the prompt update process led to modest increases in accuracy and F1\u2014although improvements were largely driven by changes in precision. While the magnitude of improvement was generally small, researchers experiencing subpar LLM annotation performance can use human-in-the-loop prompt refinement to ensure that their instructions are not the cause of poor performance. Because the improvements are small, however, it is unlikely that prompt tuning will completely resolve poor automated annotation performance at this time. Either way, a human-centered framework will remain essential: Even with improved LLM technology and automated prompt optimization strategies (Khattab et al. 2023 ###reference_b27###), researchers will always require human-labeled ground truth text data to optimize the instructions." + }, + { + "section_id": "3.2", + "parent_section_id": "3", + "section_name": "Exploiting Uncertainty in LLM Labels", + "text": "LLMs may be an effective way to identify text samples that cannot be clearly classified into the annotation categories specified by the prompt instructions. These \u201cedge-case\u201d annotations can then be manually reviewed by humans.\nOne approach is to exploit the generative LLMs\u2019 predicted token sampling process to identify an LLM\u2019s \u201cconfidence\u201d in an annotation, measured by the consistency with which it would make the annotation if asked repeatedly. By introducing some degree of randomness in the LLM sampling process through the temperature hyperparameter and by repeating an annotation task on the same text sample, we generate an empirical measure of uncertainty in the annotation label that we deem a \u201cconsistency score.\u201d141414Accessing token log probabilities, once available through the OpenAI API, may be an effective way to do the same type of selection approach. Given a vector of annotations, , with length for a given annotation task, consistency is measured as the proportion of annotations that match the modal annotation category:\nFor the analyses conducted here, we annotate every text sample at least five times at a temperature of 0.7. Put otherwise, whether GPT-4 consistently labeled the same category across all iterations. As shown in Figure 4, annotations with a consistency of 1.0 show significantly higher accuracy (19.4% increase), true positive rate (16.4% increase), and true negative rate (21.4% increase) compared to annotations with a consistency less than 1.0. Roughly 85% of annotations had a consistency of 1.0. Therefore, consistency scores offer a useful way of identifying edge cases or more difficult annotations that may require additional human inspection." + }, + { + "section_id": "3.3", + "parent_section_id": "3", + "section_name": "Manual Temperature Optimization", + "text": "Temperature, as discussed above, affects the degree of randomness introduced into LLMs\u2019 sampling process. One final way to improve LLM annotation performance may be to fine-tune this hyperparameter. Here, we report temperature fine-tuning experiments for seven of the 27 annotation tasks. In short, we annotated the same text samples numerous times using different temperatures and then compared performance across each of the temperature settings. As Figure 5 ###reference_### shows, there is not a clear relationship between performance and temperature choice, which suggests that temperature does not play a significant role in annotation performance.\n###figure_5###" + }, + { + "section_id": "3.4", + "parent_section_id": "3", + "section_name": "GPT-4 Performance Over Time", + "text": "How does LLM performance change over time? One concerning possibility is that LLMs drift in their capabilities as they are updated over time. Some research, for example, claims that GPT-4 performance has significantly decreased since its initial release (Chen, Zaharia, and Zou 2023 ###reference_b6###). This possibility is an additional reason for extra caution when deploying automated annotation tools without human validation.\nTo explore this possibility, we replicated a subset (14 of the 27 tasks) of our annotation tasks in both April 2023 and again in November 2023. Our results, displayed in the Appendix Figure A1, do not indicate meaningful changes in GPT-4 performance over time. If anything, Figure A1 shows a slight increase in performance since our initial experiments. Across the 14 tasks re-analyzed, accuracy increased by 0.007 and F1 increased by 0.022 when the same annotation procedures were conducted in November 2023. Despite these minimal changes, we cannot rule out the possibility that these models will change in the future, which again underscores the importance of a human-centered framework.\n###figure_6###" + }, + { + "section_id": "3.5", + "parent_section_id": "3", + "section_name": "Comparing GPT-4 to supervised model baseline", + "text": "We have shown that there are various circumstances where automated annotation with GPT-4 fails to align with human annotations. As a result, researchers should validate automated annotation quality against human-generated validation labels. Because automated annotation always requires some number of human validation text samples, it may be preferable in some cases to train a supervised classifier rather than simply using GPT-4 as a few-shot classification problem. Although supervised LLM classifiers (e.g., BERT) require data for fine-tuning, they are cheaper and may have similar performance to GPT-4.\nTo understand how automated annotation with GPT-4 compares to a supervised LLM classifier baseline, we replicated all 27 annotation tasks with BERT classifiers fine-tuned with either 250 or 1000 training samples.151515We are only able to replicate 14 of the 27 tasks with 1,000 training samples because the remaining nine studies had an insufficient number of human-labeled data. Figure A4 ###reference_### reports the results of these analyses. As Figure A4 ###reference_### shows, GPT-4 outperforms BERT in 18 of the 27 tasks when there are only 250 training samples to fine-tune BERT. This is likely due to the fact that BERT is unable to learn an adequate representation of the classes with few samples, especially in the presence of class imbalance. When there are minimal training samples, GPT-4 may be preferable to a supervised classifier. Nevertheless, BERT surpasses GPT-4 performance as more training data are added to the model. Specifically, BERT performs better than GPT-4 on 10 of the 14 replicated tasks when there are 1,000 training samples used to fine-tune BERT. Researchers, especially those with cost restraints, should still consider the feasibility of using supervised classification for automated annotation." + }, + { + "section_id": "4", + "parent_section_id": null, + "section_name": "Discussion", + "text": "Given their speed, relatively low cost, and ease of application, it is tempting to deploy LLMs for automated annotation without any comparison to humans. We caution against this approach in CSS and other research areas. If practitioners deploy LLMs for automated annotation without validating performance, they risk inaccuracies or pervasive bias. Other domains, like healthcare, have already documented concerning instances of bias in the deployment of LLMs. Zack et al. (2024 ###reference_b55###), for example, find that GPT-4 in applied clinical and medical applications, \u201ccan propagate, or even amplify, harmful societal biases, raising concerns about the use of GPT-4 for clinical decision support.\u201d Other domains, like algorithmic fairness (Felkner, Thompson, and May 2024 ###reference_b15###), emphasize similar types of concerns in automated annotation with LLMs. As a result, researchers across domains must ground generative AI in human judgment by validating against labels that have been prepared by annotators trained to anticipate and avoid potential biases.\nBy comparing performance across tasks and modeling strategies, we offer several advancements to the emerging literature on automated annotation using generative LLMs. First, we find that LLMs can offer high-quality labels on a wide variety of annotation tasks. Yet, automated annotation performance is subpar on a large number of tasks, which reinforces the necessity of creating a human-labeled validation set for each new annotation task. These types of procedures can help to test for convergent validity (Jacobs and Wallach 2021 ###reference_b23###). Second, recall is higher than precision on the vast majority of replicated annotation tasks. Third, various techniques, such as prompt optimization, do not improve performance enough to overcome concerns of poor performance. Fourth, while GPT-4 outperforms supervised LLM classifiers when there are minimal training samples, supervised classifiers quickly surpass GPT-4\u2019s performance as additional training data are added to the supervised model.\nIn our analyses, we find significant heterogeneity in LLM performance across a range of CSS annotation tasks. Although we are hesitant to suggest that LLMs are better at certain types of tasks than others, our results indicate that simple, summarization tasks are currently better suited for automated annotation rather than nuanced tasks that require cultural context. Because of the idiosyncrasies associated with specific tasks and datasets, however, there are circumstances where LLMs fail to deliver accurate results due to ineffective prompts, noisy text data, or difficult annotation tasks. Because these challenges will persist even as LLM technology improves,161616Even as LLM technology increases in sophistication, humans could still provide ambiguous prompt instructions, which could negatively affect LLM performance on an annotation task. researchers using LLMs for annotation must always validate on a task-by-task basis. Rigorous validation is necessary to help researchers craft effective prompts and determine whether LLM annotation is viable for their annotation tasks and datasets.\nFuture work should investigate how to effectively evaluate automated annotation with LLMs when human validation is difficult. Human annotation is especially challenging for conceptual categories that are nuanced, contested, and subjective. Prior work has shown the challenges with human coding efforts (Mikhaylov, Laver, and Benoit 2012 ###reference_b37###), which underscores the necessity of first validating human labels through inter-coder agreement metrics (Artstein and Poesio 2008 ###reference_b2###). In some cases, it may also be appropriate to model uncertainty in the human labels (Davani, D\u00edaz, and Prabhakaran 2022 ###reference_b9###; Keith et al. 2020 ###reference_b26###). Across these situations, researchers should adopt language from measurement modeling frameworks to guide their evaluation approaches (Jacobs and Wallach 2021 ###reference_b23###). However, if humans cannot effectively generate validation data then there may be concerns whether a statistical model can reasonably complete the same task either. Foregoing validation against human labels assumes that LLM judgment on an annotation task matches human judgment. The results of this study demonstrate that such an assumption is a serious concern in practice." + }, + { + "section_id": "5", + "parent_section_id": null, + "section_name": "Limitations", + "text": "We discuss several limitations of our analysis. First, our analyses primarily focus on closed-source models, although Appendix D briefly investigates an open-source alternative. While future work should continue to explore open-source automated annotations, the main thrust of our argument and analysis does not change. Either way, automated annotation should always be human-centered in some capacity.\nSecond, considering human labels as ground truth is an additional possible limitation. While our annotation sources deployed experts or multiple humans, it is not impossible that these human annotators made correlated errors. Thus, some disagreements between human labels and GPT-4 annotations may come from human error. These errors would bias performance downward. This is part of the reason we compared our results to a supervised classifier as well, since both of these approaches would suffer from ground truth errors. Providing a comparison between these two approaches allows for performance comparison with any ground truth errors in the data. In Appendix E, we explore the correlation between the human label quality (i.e., intercoder agreement) and performance across tasks.\nFor our analyses, however, the replicated annotation tasks and data come from peer-reviewed research in high-impact journals. We believe that this helps to reduce concerns about data annotation quality. The annotation procedures in these tasks were approved by IRB protocols and were scrutinized by independent reviewers for publication in a high-impact journal. Nevertheless, it is critical to acknowledge that researchers should collect high-quality human labels and validate them as well (Artstein and Poesio 2008 ###reference_b2###)." + }, + { + "section_id": "6", + "parent_section_id": null, + "section_name": "Ethics Statement", + "text": "In addition to adding the mandatory Paper Checklist (added after all appendices), we also include a brief ethics statement. Our research positively contributes to society and human well-being without violating any social contracts. We test generative AI tools that can help computational social scientists investigate the social world. Using the methods we explore will help researchers better understand a wide range of challenging social problems. Moreover, because our analyses require significantly fewer resources than alternatives, we believe that our results can help reduce inequities in resources across scientists.\nBecause of the potential negative impacts of deploying biased models, we stress the importance of human validation throughout our paper. We argue that it is critical to develop rigorous evaluation standards that are human-centered. This is part of the reason why we centered our experiments on data sets and tasks less prone to contamination.\nWe also take steps to reduce potential negative impacts of this research. Specifically, we respected confidentiality and privacy issues. We also adhered to each data repository\u2019s usage and replication policies. Moreover, all of the original studies received IRB approval and our experiments conformed to the same safety protocols. All data was anonymized by the original publication authors. Appendix B provides more details on human annotation protocols, which were conducted by the original authors and received IRB approval." + } + ], + "appendix": [ + { + "section_id": "Appendix 1", + "parent_section_id": null, + "section_name": "Appendix A Appendix A: Replication data and tasks", + "text": "" + }, + { + "section_id": "Appendix 2", + "parent_section_id": null, + "section_name": "Appendix B Appendix B: Human labels in replication data", + "text": "We use a novel dataset for this analysis, compiled using human-labeled data from recent studies in high-impact journals (see A1 ###reference_###). Rather than attempting to assess the quality of the annotation procedure used by the original authors, we rely on the peer-reviewed publication process as an indication that these studies are at least of representative quality of computational social science research.\nThe human-labeled classification tasks in these studies are ones that can be tested without concerns of contamination. Although the raw text data (e.g., Facebook posts and tweets) are possibly included the LLM pretraining data, the associated labels from the original annotators are certainly not in the pretraining data. This is because the human-generated labels accompanying each text sample (e.g., whether a specific tweet referenced a gender identity frame) are not publicly available. Because the text with the human labels is not included in the pretraining data, there is minimal cause for contamination concerns.\nTo replicate their annotation procedures, we use verbatim (or, if the original materials are not available, paraphrase) the annotation instructions used in the original study for our initial prompts. All prompt instructions are included in the supplementary material. The human coders for these datasets were experts, undergraduate RAs, or crowd-sourced workers. The original authors had received IRB approval for each study we replicate. Further details about the annotation procedures for each replicated study can be found in the methodological details for those studies." + }, + { + "section_id": "Appendix 3", + "parent_section_id": null, + "section_name": "Appendix C Appendix C: Methodological details", + "text": "To establish a baseline for classification performance, we fine-tuned BERT classifiers with varying training data strategies: n=250 human samples and n=1,000 human samples. We use BERT as it is among the most popular models used for supervised classification in computational social science. Because the hyperparameters that yield strongest performance may differ across tasks, we used a grid search. For each classifier, we searched over 18 combinations of learning rate (1e-5, 2e-5, and 5e-5), batch size (8 and 16), and number of epochs (2, 4, 6). This grid search was conducted on a subsample of 250 text samples per task. We identified the combination of hyperparameters with the highest F1 and retained that combination for all models for that particular task (see best-performing hyperparameter configurations in Table A3 ###reference_###). Hyperparameters that remained constant across tasks and details about model architecture are shown in Table A4 ###reference_###. We evaluate each model on 1,000 held-out text samples with labels from human annotators.\nThis analysis was conducted in Python 3.10.12 using HuggingFace\u2019s Transformers (Wolf et al. 2020 ###reference_b53###) library and PyTorch libraries (Paszke et al. 2019 ###reference_b42###). For data preprocessing, we use Python Pandas (McKinney 2011 ###reference_b35###). We used Google Colab as our computing environment, totalling 215 T4 GPU compute units (approximately 421.4 GPU hours). In the supplementary material, we include all code to run our supervised training procedures." + }, + { + "section_id": "Appendix 4", + "parent_section_id": null, + "section_name": "Appendix D Appendix D: Comparing to an open-source alternative", + "text": "Due to various concerns about closed-source models (e.g., Spirling, 2023), we also generate few-shot labels using Mistral-7B from the same procedures employed in the GPT-4 few-shot model. We include the Mistral-7B few-shot model as a smaller, open-source alternative to GPT-4. Mistral-7B was selected because the model weights are available for download and it has higher performance than Llama-13B (Jiang et al. 2023 ###reference_b24###).\nFigure A2 ###reference_### compares automated annotation classification performance between GPT-4 and Mistral-7B. The evaluation sets for these experiments are the same as the main analyses previously conducted. Figure A2 ###reference_### shows that there is a fairly sizeable gap between the open-source (Mistral-7B) and closed-source (GPT-4) few-shot models. F1 scores for Mistral-7B are 0.16 worse, on average, than GPT-4, although this may be expected from a significantly smaller and free-to-use model. Cheaper, smaller open models may perform worse on automated annotation than their larger, propriety counterparts. Thus, these findings further reinforce the necessity of human validation to confirm annotation quality." + }, + { + "section_id": "Appendix 5", + "parent_section_id": null, + "section_name": "Appendix E Appendix E: Relationship between task difficulty and performance", + "text": "Here, we investigate automated annotation performance across different text and task difficulties. Figure A3 ###reference_### measures text and task complexity in different ways and compares difficulty to GPT-4 annotation performance. The top two panes in Figure A3 ###reference_### show that text length and text readability (Flesch-Kincaid grade level score)171717The Flesch-Kincaid Grade Level metric signals how challenging a text passage is for a human to read. The numeric score translates to the recommended grade level required to read the text. The score is calculated based on the average number of words per sentence and average number of syllables per word. are uncorrelated to automated annotation performance. Both of these measures were calculated on the text samples in the original studies. Neither of these measures of text difficulty predict poor performance.\nThe bottom pane of Figure A3 ###reference_### shows the correlation between intercoder reliability and GPT-4 F1 performance. Across the 27 replicated tasks, 21 tasks in the original studies reported metrics on intercoder reliability. 16 studies utilized Krippendorff\u2019s alpha, while five studies calculated Cohen\u2019s kappa. If the original study involved a multi-class task, a single reliability metric was noted in the study, pooled across each label. Figure A3 ###reference_### illustrates that, when intercoder reliability is low, GPT-4 also has low performance. This could possibly be due to inadequate conceptualization of the annotation category, which might lead to inconsistent guidelines for sample categorization. Categories with medium to high intercoder reliability display higher performance than the categories with low intercoder reliability. Beyond this, however, there is minimal correlation between high intercoder reliability and strong GPT performance. Thus, while it is possible that poor performance was the result of difficult coding tasks, the evidence provided here suggests that strong intercoder reliability only weakly predicts strong performance." + }, + { + "section_id": "Appendix 6", + "parent_section_id": null, + "section_name": "Appendix F Appendix F: Miscellaneous additional information", + "text": "Additional sources:\nRobot image (used in Figure 1): https://commons.wikimedia.org/wiki/File:Grey\u02d9cartoon\u02d9robot.png\nHuman silhouette image (used in Figure 1): https://commons.wikimedia.org/wiki/File:SVG\u02d9Human\u02d9Silhouette.svg\n###figure_7### ###figure_8### ###figure_9### ###figure_10###" + }, + { + "section_id": "Appendix 7", + "parent_section_id": null, + "section_name": "Appendix G Checklist", + "text": "For most authors\u2026\nWould answering this research question advance science without violating social contracts, such as violating privacy norms, perpetuating unfair profiling, exacerbating the socio-economic divide, or implying disrespect to societies or cultures?\nYes, see Ethics Statement\nDo your main claims in the abstract and introduction accurately reflect the paper\u2019s contributions and scope?\nYes, see Results\nDo you clarify how the proposed methodological approach is appropriate for the claims made?\nYes, see Replication Procedures\nDo you clarify what are possible artifacts in the data used, given population-specific distributions?\nNA\nDid you describe the limitations of your work?\nYes, see Limitations\nDid you discuss any potential negative societal impacts of your work?\nYes, see Ethics Statement\nDid you discuss any potential misuse of your work?\nYes, see Ethics Statement\nDid you describe steps taken to prevent or mitigate potential negative outcomes of the research, such as data and model documentation, data anonymization, responsible release, access control, and the reproducibility of findings?\nYes, see Ethics Statement\nHave you read the ethics review guidelines and ensured that your paper conforms to them?\nYes\nAdditionally, if your study involves hypotheses testing\u2026\nDid you clearly state the assumptions underlying all theoretical results?\nNA\nHave you provided justifications for all theoretical results?\nNA\nDid you discuss competing hypotheses or theories that might challenge or complement your theoretical results?\nNA\nHave you considered alternative mechanisms or explanations that might account for the same outcomes observed in your study?\nNA\nDid you address potential biases or limitations in your theoretical framework?\nNA\nHave you related your theoretical results to the existing literature in social science?\nNA\nDid you discuss the implications of your theoretical results for policy, practice, or further research in the social science domain?\nNA\nAdditionally, if you are including theoretical proofs\u2026\nDid you state the full set of assumptions of all theoretical results?\nNA\nDid you include complete proofs of all theoretical results?\nNA\nAdditionally, if you ran machine learning experiments\u2026\nDid you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)?\nYes, attached as supplementary materials\nDid you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?\nYes, see Appendix C\nDid you report error bars (e.g., with respect to the random seed after running experiments multiple times)?\nNA\nDid you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)?\nYes, see Appendix C\nDo you justify how the proposed evaluation is sufficient and appropriate to the claims made?\nYes, see Introduction, Results\nDo you discuss what is \u201cthe cost\u201c of misclassification and fault (in)tolerance?\nYes, see Introduction, Discussion\nAdditionally, if you are using existing assets (e.g., code, data, models) or curating/releasing new assets, without compromising anonymity\u2026\nIf your work uses existing assets, did you cite the creators?\nYes\nDid you mention the license of the assets?\nNA\nDid you include any new assets in the supplemental material or as a URL?\nNA\nDid you discuss whether and how consent was obtained from people whose data you\u2019re using/curating?\nYes, see Appendix B\nDid you discuss whether the data you are using/curating contains personally identifiable information or offensive content?\nYes, see Ethics Statement and Appendix B\nIf you are curating or releasing new datasets, did you discuss how you intend to make your datasets FAIR?\nNA\nIf you are curating or releasing new datasets, did you create a Datasheet for the Dataset (see Gebru et al. (2021 ###reference_b16###))?\nNA\nAdditionally, if you used crowdsourcing or conducted research with human subjects, without compromising anonymity\u2026\nDid you include the full text of instructions given to participants and screenshots?\nNo, because we do not work directly with human subjects from the original study. In the supplementary materials, we include LLM prompt instructions that are based on instructions given to humans.\nDid you describe any potential participant risks, with mentions of Institutional Review Board (IRB) approvals?\nNo, because we do not work directly with human subjects from the original study. All replicated studies received IRB approval.\nDid you include the estimated hourly wage paid to participants and the total amount spent on participant compensation?\nNo, because we do not work directly with human subjects from the original study.\nDid you discuss how data is stored, shared, and deidentified?\nNo, because we are not publicly releasing the data without the original authors permission. All data was already deidentified." + } + ], + "tables": { + "1": { + "table_html": "
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
MetricMinimumMeanMedianMaximum
F10.060.660.710.97
Accuracy0.670.860.850.98
Precision0.030.620.650.96
Recall0.250.750.830.98
\n
Table 1: Automated annotation performance across 27 tasks from 11 datasets, assessed on 1000 human-generated labels.
\n
", + "capture": "Table 1: Automated annotation performance across 27 tasks from 11 datasets, assessed on 1000 human-generated labels." + }, + "2": { + "table_html": "
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n\nAuthor(s)\n\n\n\nTitle\n\n\n\nJournal\n\n\n\nYear\n\n
\n\nBusby, and Gubler, Hawkins\n\n\n\nFraming and blame attribution in populist rhetoric\n\n\n\nJournal of Politics\n\n\n\n2019\n\n
\n\nCard et al.\n\n\n\nComputational analysis of 140 years of US political speeches reveals more positive but increasingly polarized framing of immigration\n\n\n\nPNAS\n\n\n\n2022\n\n
\n\nCusimano and Goodwin\n\n\n\nPeople judge others to have more voluntary control over beliefs than they themselves do\n\n\n\nJournal of Personality and Social Psychology\n\n\n\n2020\n\n
\n\nGohdes\n\n\n\nRepression Technology: Internet Accessibility and State Violence\n\n\n\nAmerican Journal of Political Science\n\n\n\n2020\n\n
\n\nHopkins, Lelkes, and Wolken\n\n\n\nThe Rise of and Demand for Identity-Oriented Media Coverage\n\n\n\nAmerican Journal of Political Science\n\n\n\n2024\n\n
\n\nM\u00fcller\n\n\n\nThe Temporal Focus of Campaign Communication\n\n\n\nJournal of Politics\n\n\n\n2021\n\n
\n\nPeng, Romero, and Horvat\n\n\n\nDynamics of cross-platform attention to retracted papers\n\n\n\nPNAS\n\n\n\n2022\n\n
\n\nSaha et al.\n\n\n\nOn the rise of fear speech in online social media\n\n\n\nPNAS\n\n\n\n2022\n\n
\n\nSchub\n\n\n\nInforming the Leader: Bureaucracies and International Crises\n\n\n\nAmerican Political Science Review\n\n\n\n2022\n\n
\n\nWojcieszak et al.\n\n\n\nMost users do not follow political elites on Twitter; those who do show overwhelming preferences for ideological congruity\n\n\n\nScience Advances\n\n\n\n2022\n\n
\n\nYu and Zhang\n\n\n\nThe Impact of Social Identity Conflict on Planning Horizons\n\n\n\nJournal of Personality and Social Psychology\n\n\n\n2022\n\n
\n
Table A1: Sources of annotation tasks replicated in analysis.
\n
", + "capture": "Table A1: Sources of annotation tasks replicated in analysis." + }, + "3": { + "table_html": "
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
StudyAnnotation tasksAnnotator(s)
\n\nBusby, Gubler, & Hawkins (2019)\n\n\n\nLabel open-ended text from participants in experiment for whether they (1) attribute blame to a specific actor, (2) attribute blame to a nefarious elite actor, or (3) include a positive mention of the collective people\n\n\n\nResearch assistant\n\n
\n\nCard et al. (2022)\n\n\n\nLabel congressional speeches for whether they are about immigration; also label tone as (1) proimmigration, (2) antiimmigration, or (3) neutral\n\n\n\nResearch assistant\n\n
\n\nCusimano & Goodwin (2020)\n\n\n\nLabel open-ended text on climate change from participants in experiment for the presence of (1) generic reasoning about beliefs and (2) supporting evidence for the belief\n\n\n\nResearch assistant\n\n
\n\nGohdes (2020)\n\n\n\nLabel Syrian death records based on type of killing: targeted or untargeted\n\n\n\nExpert\n\n
\n\nHopkins, Lelkes, & Wolken (2023)\n\n\n\nLabel if news content (headline, tweet, or Facebook share blurb) references social groups defined by (1) race/ethnicity; (2) gender/sexuality; (3) politics; (4) religion\n\n\n\nCrowd\n\n
\n\nM\u00fcller (2021)\n\n\n\nLabel sampled sentences from political party manifestos for if the temporal direction is (1) past, (2) present, or (3) future\n\n\n\nExpert\n\n
\n\nPeng, Romero, & Horvat (2022)\n\n\n\nLabel tweets for whether they criticize academic papers\n\n\n\nExpert\n\n
\n\nSaha et al. (2020)\n\n\n\nLabel Gab posts for whether they include (1) fear speech and/or (2) hate speech\n\n\n\nCrowd\n\n
\n\nSchub (2020)\n\n\n\nLabel texts from Cold War crises as political or military\n\n\n\nExpert\n\n
\n\nWojcieszak et al. (2022)\n\n\n\nLabel a quote tweet as (1) negative, (2) neutral or (3) positive toward the message and/or the political actor being quoted, independently of the tone of the original message\n\n\n\nExpert\n\n
\n\nYu & Zhang (2023)\n\n\n\nLabel open-ended text about plans for the future from participants in experiment for whether they are about the (1) proximate future or (2) distant future\n\n\n\nResearch assistant\n\n
\n
Table A2: Descriptions of annotation tasks replicated in analysis.
\n
", + "capture": "Table A2: Descriptions of annotation tasks replicated in analysis." + }, + "4": { + "table_html": "
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n\nStudy\n\n\n\nTask\n\n\n\nHyperparameters\n\n
\n\nBusby\n\n\n\nClassify specific blame\n\n\n\nlearning rate (5e-05), batch size (8), epochs (6)\n\n
\n\nClassify elite blame\n\n\n\nlearning rate (5e-05), batch size (8), epochs (4)\n\n
\n\nClassify collective positive\n\n\n\nlearning rate (5e-05), batch size (8), epochs (6)\n\n
\n\nCusimano et al.\n\n\n\nClassify generic belief\n\n\n\nlearning rate (5e-05), batch size (8), epochs (6)\n\n
\n\nClassify evidence belief\n\n\n\nlearning rate (2e-05), batch size (16), epochs (2)\n\n
\n\nCard et al.\n\n\n\nClassify immigration speeches\n\n\n\nlearning rate (5e-05), batch size (8), epochs (4)\n\n
\n\nClassify pro-immigration speeches\n\n\n\nlearning rate (5e-05), batch size (16), epochs (6)\n\n
\n\nClassify anti-immigration speeches\n\n\n\nlearning rate (5e-05), batch size (8), epochs (6)\n\n
\n\nClassify neutral immigration speeches\n\n\n\nlearning rate (5e-05), batch size (8), epochs (4)\n\n
\n\nGohdes\n\n\n\nClassify targeted killings\n\n\n\nlearning rate (5e-05), batch size (8), epochs (4)\n\n
\n\nClassify untargeted killings\n\n\n\nlearning rate (5e-05), batch size (16), epochs (6)\n\n
\n\nHopkins et al.\n\n\n\nClassify race/ethnicity\n\n\n\nlearning rate (2e-05), batch size (8), epochs (4)\n\n
\n\nClassify gender\n\n\n\nlearning rate (5e-05), batch size (8), epochs (6)\n\n
\n\nClassify political groups\n\n\n\nlearning rate (5e-05), batch size (16), epochs (6)\n\n
\n\nClassify religious groups\n\n\n\nlearning rate (5e-05), batch size (8), epochs (6)\n\n
\n\nM\u00fcller\n\n\n\nClassify past\n\n\n\nlearning rate (5e-05), batch size (8), epochs (4)\n\n
\n\nClassify present\n\n\n\nlearning rate (5e-05), batch size (8), epochs (4)\n\n
\n\nClassify future\n\n\n\nlearning rate (2e-05), batch size (8), epochs (6)\n\n
\n\nPeng et al.\n\n\n\nClassify criticism\n\n\n\nlearning rate (5e-05), batch size (8), epochs (6)\n\n
\n\nSaha et al.\n\n\n\nClassify fear speech\n\n\n\nlearning rate (5e-05), batch size (8), epochs (6)\n\n
\n\nClassify hate speech\n\n\n\nlearning rate (5e-05), batch size (8), epochs (4)\n\n
\n\nSchub\n\n\n\nClassify political or military text\n\n\n\nlearning rate (5e-05), batch size (8), epochs (6)\n\n
\n\nWojcieszak et al.\n\n\n\nClassify positive\n\n\n\nlearning rate (2e-05), batch size (8), epochs (6)\n\n
\n\nClassify negative\n\n\n\nlearning rate (5e-05), batch size (16), epochs (6)\n\n
\n\nClassify neutral\n\n\n\nlearning rate (2e-05), batch size (8), epochs (2)\n\n
\n\nYu et al.\n\n\n\nClassify proximate future\n\n\n\nlearning rate (5e-05), batch size (8), epochs (6)\n\n
\n\nClassify distant future\n\n\n\nlearning rate (5e-05), batch size (8), epochs (6)\n\n
\n
Table A3: Hyperparameter settings per task.
\n
", + "capture": "Table A3: Hyperparameter settings per task." + }, + "5": { + "table_html": "
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n\nBERT-base\n\n
\n\n# parameters\n\n\n\n110m\n\n
\n\n# attention heads\n\n\n\n12\n\n
\n\nHidden dim.\n\n\n\n768\n\n
\n\nFeedforward dim.\n\n\n\n3072\n\n
\n\nActivation\n\n\n\nGELU\n\n
\n\nDropout\n\n\n\n0.1 1\n\n
\n\nOptimizer\n\n\n\nAdam\n\n
\n\nWeight decay\n\n\n\n0.01\n\n
\n
Table A4: Model architectures and additional hyperparameters.
\n
", + "capture": "Table A4: Model architectures and additional hyperparameters." + } + }, + "image_paths": { + "1(a)": { + "figure_path": "2409.09467v2_figure_1(a).png", + "caption": "Figure 1: Human-in-the-loop workflow for augmenting text annotation with generative LLMs", + "url": "http://arxiv.org/html/2409.09467v2/extracted/5870309/human.png" + }, + "1(b)": { + "figure_path": "2409.09467v2_figure_1(b).png", + "caption": "Figure 1: Human-in-the-loop workflow for augmenting text annotation with generative LLMs", + "url": "http://arxiv.org/html/2409.09467v2/extracted/5870309/robot.png" + }, + "2": { + "figure_path": "2409.09467v2_figure_2.png", + "caption": "Figure 2: Precision and recall for 27 annotation tasks compared to human labels. Points sharing the same color are conducted on the same text data set.", + "url": "http://arxiv.org/html/2409.09467v2/x1.png" + }, + "3": { + "figure_path": "2409.09467v2_figure_3.png", + "caption": "Figure 3: Change in LLM annotation performance on training data after one round of prompt updates.", + "url": "http://arxiv.org/html/2409.09467v2/x2.png" + }, + "4": { + "figure_path": "2409.09467v2_figure_4.png", + "caption": "Figure 4: Relationship between consistency score and accuracy, TPR, and TNR. Lines are linear trend trends weighted by the number of samples that fall under each density score.", + "url": "http://arxiv.org/html/2409.09467v2/x3.png" + }, + "5": { + "figure_path": "2409.09467v2_figure_5.png", + "caption": "Figure 5: Relationship between temperature and F1 score.", + "url": "http://arxiv.org/html/2409.09467v2/x4.png" + }, + "6": { + "figure_path": "2409.09467v2_figure_6.png", + "caption": "Figure A1: Examining GPT-4 performance over time", + "url": "http://arxiv.org/html/2409.09467v2/x5.png" + }, + "7": { + "figure_path": "2409.09467v2_figure_7.png", + "caption": "Figure A2: Comparing GPT-4 and Mistral-7B few-shot classifiers", + "url": "http://arxiv.org/html/2409.09467v2/x6.png" + }, + "8": { + "figure_path": "2409.09467v2_figure_8.png", + "caption": "Figure A3: Relationship between text/task difficulty and performance.", + "url": "http://arxiv.org/html/2409.09467v2/x7.png" + }, + "9": { + "figure_path": "2409.09467v2_figure_9.png", + "caption": "Figure A4: Supervised text classifiers outperform GPT-4 with additional training samples.", + "url": "http://arxiv.org/html/2409.09467v2/x8.png" + } + }, + "validation": true, + "references": [ + { + "1": { + "title": "Reducing multiclass to binary: A unifying approach for margin classifiers.", + "author": "Allwein, E. L.; Schapire, R. E.; and Singer, Y. 2000.", + "venue": "Journal of machine learning research, 1(Dec): 113\u2013141.", + "url": null + } + }, + { + "2": { + "title": "Inter-coder agreement for computational linguistics.", + "author": "Artstein, R.; and Poesio, M. 2008.", + "venue": "Computational linguistics, 34(4): 555\u2013596.", + "url": null + } + }, + { + "3": { + "title": "Using Large Language Models for Qualitative Analysis can Introduce Serious Bias.", + "author": "Ashwin, J.; Chhabra, A.; and Rao, V. 2023.", + "venue": "arXiv preprint arXiv:2309.17147.", + "url": null + } + }, + { + "4": { + "title": "Language Models are Few-Shot Learners.", + "author": "Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J. D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; Agarwal, S.; Herbert-Voss, A.; Krueger, G.; Henighan, T.; Child, R.; Ramesh, A.; Ziegler, D.; Wu, J.; Winter, C.; Hesse, C.; Chen, M.; Sigler, E.; Litwin, M.; Gray, S.; Chess, B.; Clark, J.; Berner, C.; McCandlish, S.; Radford, A.; Sutskever, I.; and Amodei, D. 2020.", + "venue": "In Larochelle, H.; Ranzato, M.; Hadsell, R.; Balcan, M.; and Lin, H., eds., Advances in Neural Information Processing Systems, volume 33, 1877\u20131901. Curran Associates, Inc.", + "url": null + } + }, + { + "5": { + "title": "Computational analysis of 140 years of US political speeches reveals more positive but increasingly polarized framing of immigration.", + "author": "Card, D.; Chang, S.; Becker, C.; Mendelsohn, J.; Voigt, R.; Boustan, L.; Abramitzky, R.; and Jurafsky, D. 2022.", + "venue": "Proceedings of the National Academy of Sciences of the United States of America, 31.", + "url": null + } + }, + { + "6": { + "title": "How is ChatGPT\u2019s behavior changing over time?", + "author": "Chen, L.; Zaharia, M.; and Zou, J. 2023.", + "venue": "arXiv:2307.09009.", + "url": null + } + }, + { + "7": { + "title": "Can Large Language Models Be an Alternative to Human Evaluations?", + "author": "Chiang, C.-H.; and Lee, H.-y. 2023.", + "venue": "arXiv preprint arXiv:2305.01937.", + "url": null + } + }, + { + "8": { + "title": "People judge others to have more voluntary control over beliefs than they themselves do.", + "author": "Cusimano, C.; and Goodwin, G. P. 2020.", + "venue": "Journal of Personality and Social Psychology, 119.", + "url": null + } + }, + { + "9": { + "title": "Dealing with disagreements: Looking beyond the majority vote in subjective annotations.", + "author": "Davani, A. M.; D\u00edaz, M.; and Prabhakaran, V. 2022.", + "venue": "Transactions of the Association for Computational Linguistics, 10: 92\u2013110.", + "url": null + } + }, + { + "10": { + "title": "Bert: Pre-training of deep bidirectional transformers for language understanding.", + "author": "Devlin, J.; Chang, M.-W.; Le, K.; and Toutanova, K. 2019.", + "venue": "In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), 4171\u20134186.", + "url": null + } + }, + { + "11": { + "title": "Is GPT-3 a Good Data Annotator?", + "author": "Ding, B.; Qin, C.; Liu, L.; Bing, L.; Joty, S.; and Li, B. 2022.", + "venue": "arXiv:2212.10450.", + "url": null + } + }, + { + "12": { + "title": "How to make causal inferences using texts.", + "author": "Egami, N.; Fong, C. J.; Grimmer, J.; Roberts, M. E.; and Stewart, B. M. 2022.", + "venue": "Science Advances, 8(42): eabg2652.", + "url": null + } + }, + { + "13": { + "title": "Use of large language models as a scalable approach to understanding public health discourse.", + "author": "Espinosa, L.; and Salath\u00e9, M. 2024.", + "venue": "medRxiv, 2024\u201302.", + "url": null + } + }, + { + "14": { + "title": "Governing AI safety through independent audits.", + "author": "Falco, G.; Shneiderman, B.; Badger, J.; Carrier, R.; Dahbura, A.; Danks, D.; Eling, M.; Goodloe, A.; Gupta, J.; Hart, C.; Jirotka, M.; Johnson, H.; LaPointe, C.; Llorens, A. J.; Mackworth, A. K.; Maple, C.; P\u00e1lsson, S. E.; Pasquale, F.; Winfield, A.; and Yeong, Z. K. 2021.", + "venue": "Nature Machine Intelligence, 7: 566\u2013571.", + "url": null + } + }, + { + "15": { + "title": "GPT is Not an Annotator: The Necessity of Human Annotation in Fairness Benchmark Construction.", + "author": "Felkner, V.; Thompson, J.; and May, J. 2024.", + "venue": "In Ku, L.-W.; Martins, A.; and Srikumar, V., eds., Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 14104\u201314115. Bangkok, Thailand: Association for Computational Linguistics.", + "url": null + } + }, + { + "16": { + "title": "Datasheets for datasets.", + "author": "Gebru, T.; Morgenstern, J.; Vecchione, B.; Vaughan, J. W.; Wallach, H.; Iii, H. D.; and Crawford, K. 2021.", + "venue": "Communications of the ACM, 64(12): 86\u201392.", + "url": null + } + }, + { + "17": { + "title": "ChatGPT outperforms crowd workers for text-annotation tasks.", + "author": "Gilardi, F.; Alizadeh, M.; and Kubli, M. 2023.", + "venue": "Proceedings of the National Academy of Sciences, 120(30): e2305016120.", + "url": null + } + }, + { + "18": { + "title": "Text as Data: A New Framework for Machine Learning and the Social Sciences.", + "author": "Grimmer, J.; Roberts, M. E.; and Stewart, B. 2022.", + "venue": "Princeton University Press.", + "url": null + } + }, + { + "19": { + "title": "Text as Data: The Promise and Pitfalls of Automatic Content Analysis Methods for Political Texts.", + "author": "Grimmer, J.; and Stewart, B. M. 2013.", + "venue": "Political Analysis, 21(3): 267\u2013297.", + "url": null + } + }, + { + "20": { + "title": "AnnoLLM: Making Large Language Models to Be Better Crowdsourced Annotators.", + "author": "He, X.; Lin, Z.; Gong, Y.; Jin, A.-L.; Zhang, H.; Lin, C.; Jiao, J.; Yiu, S. M.; Duan, N.; and Chen, W. 2023.", + "venue": "arXiv:2303.16854.", + "url": null + } + }, + { + "21": { + "title": "The Rise of and Demand for Identity-Oriented Media Coverage.", + "author": "Hopkins, D. J.; Lelkes, Y.; and Wolken, S. 2024.", + "venue": "American Journal of Political Science.", + "url": null + } + }, + { + "22": { + "title": "Is chatgpt better than human annotators? potential and limitations of chatgpt in explaining implicit hate speech.", + "author": "Huang, F.; Kwak, H.; and An, J. 2023.", + "venue": "arXiv:2302.07736.", + "url": null + } + }, + { + "23": { + "title": "Measurement and fairness.", + "author": "Jacobs, A. Z.; and Wallach, H. 2021.", + "venue": "In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, 375\u2013385.", + "url": null + } + }, + { + "24": { + "title": "Mistral 7B.", + "author": "Jiang, A. Q.; Sablayrolles, A.; Mensch, A.; Bamford, C.; Chaplot, D. S.; de las Casas, D.; Bressand, F.; Lengyel, G.; Lample, G.; Saulnier, L.; Lavaud, L. R.; Lachaux, M.-A.; Stock, P.; Scao, T. L.; Lavril, T.; Wang, T.; Lacroix, T.; and Sayed, W. E. 2023.", + "venue": "arXiv:2310.06825.", + "url": null + } + }, + { + "25": { + "title": "Aggressive sampling for multi-class to binary reduction with applications to text classification.", + "author": "Joshi, B.; Amini, M. R.; Partalas, I.; Iutzeler, F.; and Maximov, Y. 2017.", + "venue": "Advances in Neural Information Processing Systems, 30.", + "url": null + } + }, + { + "26": { + "title": "Uncertainty over uncertainty: Investigating the assumptions, annotations, and text measurements of economic policy uncertainty.", + "author": "Keith, K. A.; Teichmann, C.; O\u2019Connor, B.; and Meij, E. 2020.", + "venue": "arXiv preprint arXiv:2010.04706.", + "url": null + } + }, + { + "27": { + "title": "DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines.", + "author": "Khattab, O.; Singhvi, A.; Maheshwari, P.; Zhang, Z.; Santhanam, K.; Vardhamanan, S.; Haq, S.; Sharma, A.; Joshi, T. T.; Moazam, H.; Miller, H.; Zaharia, M.; and Potts, C. 2023.", + "venue": "arXiv:2310.03714.", + "url": null + } + }, + { + "28": { + "title": "Testing Causal Theories with Learned Proxies.", + "author": "Knox, D.; Lucas, C.; and Cho, W. K. T. 2022.", + "venue": "Annual Review of Political Science, 25.", + "url": null + } + }, + { + "29": { + "title": "Content Analysis: An Introduction to Its Methodology.", + "author": "Krippendorff, K. 2018.", + "venue": "Sage, 4 edition.", + "url": null + } + }, + { + "30": { + "title": "Harnessing Artificial Intelligence to Combat Online Hate: Exploring the Challenges and Opportunities of Large Language Models in Hate Speech Detection.", + "author": "Kumarage, T.; Bhattacharjee, A.; and Garland, J. 2024.", + "venue": "arXiv:2403.08035.", + "url": null + } + }, + { + "31": { + "title": "Task Contamination: Language Models May Not Be Few-Shot Anymore.", + "author": "Li, C.; and Flanigan, J. 2023.", + "venue": "arXiv:2312.16337.", + "url": null + } + }, + { + "32": { + "title": "CoAnnotating: Uncertainty-Guided Work Allocation between Human and Large Language Models for Data Annotation.", + "author": "Li, M.; Shi, T.; Ziems, C.; Kan, M.-Y.; Chen, N. F.; Liu, Z.; and Yang, D. 2023.", + "venue": "arXiv:2310.15638.", + "url": null + } + }, + { + "33": { + "title": "A review on the combination of binary classifiers in multiclass problems.", + "author": "Lorena, A. C.; De Carvalho, A. C.; and Gama, J. M. 2008.", + "venue": "Artificial Intelligence Review, 30: 19\u201337.", + "url": null + } + }, + { + "34": { + "title": "Data contamination: From memorization to exploitation.", + "author": "Magar, I.; and Schwartz, R. 2022.", + "venue": "arXiv:2203.08242.", + "url": null + } + }, + { + "35": { + "title": "pandas: a foundational Python library for data analysis and statistics.", + "author": "McKinney, W. 2011.", + "venue": "Python for high performance and scientific computing, 14(9): 1\u20139.", + "url": null + } + }, + { + "36": { + "title": "Does GPT-3 know what the Most Important Issue is? Using Large Language Models to Code Open-Text Social Survey Responses At Scale.", + "author": "Mellon, J.; Bailey, J.; Scott, R.; Breckwoldt, J.; and Miori, M. 2022.", + "venue": "Working paper.", + "url": null + } + }, + { + "37": { + "title": "Coder reliability and misclassification in the human coding of party manifestos.", + "author": "Mikhaylov, S.; Laver, M.; and Benoit, K. R. 2012.", + "venue": "Political Analysis, 20(1): 78\u201391.", + "url": null + } + }, + { + "38": { + "title": "The temporal focus of campaign communication.", + "author": "M\u00fcller, S. 2022.", + "venue": "The Journal of Politics, 84(1): 585\u2013590.", + "url": null + } + }, + { + "39": { + "title": "The Content Analysis Guidebook.", + "author": "Neuendorf, K. A. 2016.", + "venue": "Sage Publications.", + "url": null + } + }, + { + "40": { + "title": "ChatGPT for Text Annotation? Mind the Hype!", + "author": "Ollion, E.; Shen, R.; Macanovic, A.; and Chatelain, A. 2023.", + "venue": "arXiv:https://osf.io/preprints/socarxiv/x58kn.", + "url": null + } + }, + { + "41": { + "title": "Do the rewards justify the means? measuring trade-offs between rewards and ethical behavior in the machiavelli benchmark.", + "author": "Pan, A.; Chan, J. S.; Zou, A.; Li, N.; Basart, S.; Woodside, T.; Zhang, H.; Emmons, S.; and Hendrycks, D. 2023.", + "venue": "Proceedings of the 40th International Conference on Machine Learning, 26837\u201326867.", + "url": null + } + }, + { + "42": { + "title": "Pytorch: An imperative style, high-performance deep learning library.", + "author": "Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; and et al., T. K. 2019.", + "venue": "Advances in neural information processing systems.", + "url": null + } + }, + { + "43": { + "title": "Dynamics of cross-platform attention to retracted papers.", + "author": "Peng, H.; Romero, D. M.; and Horvat, E.-A. 2022.", + "venue": "Proceedings of the National Academy of Sciences, 119(25): 585\u2013590.", + "url": null + } + }, + { + "44": { + "title": "Testing the Reliability of ChatGPT for Text Annotation and Classification: A Cautionary Remark.", + "author": "Reiss, M. V. 2023.", + "venue": "arXiv:2304.11085.", + "url": null + } + }, + { + "45": { + "title": "Adjusting for Confounding with Text Matching.", + "author": "Roberts, M. E.; Stewart, B. M.; and Nielsen, R. A. 2020.", + "venue": "American Journal of Political Science, 64(4): 887\u2013903.", + "url": null + } + }, + { + "46": { + "title": "Informing the Leader: Bureaucracies and International Crises.", + "author": "Schub, R. 2022.", + "venue": "American Political Science Review, 116.", + "url": null + } + }, + { + "47": { + "title": "Human-centered artificial intelligence: Reliable, safe and trustworthy.", + "author": "Shneiderman, B. 2020.", + "venue": "International Journal of Human\u2013Computer Interaction, 6: 495\u2013504.", + "url": null + } + }, + { + "48": { + "title": "Why open-source generative AI models are an ethical way forward for science.", + "author": "Spirling, A. 2023.", + "venue": "Nature, 413.", + "url": null + } + }, + { + "49": { + "title": "From humans to machines: can ChatGPT-like LLMs effectively replace human annotators in NLP tasks.", + "author": "Thapa, S.; Naseem, U.; and Nasim, M. 2023.", + "venue": "Workshop Proceedings of the 17th International AAAI Conference on Web and Social Media.", + "url": null + } + }, + { + "50": { + "title": "ChatGPT-4 Outperforms Experts and Crowd Workers in Annotating Political Twitter Messages with Zero-Shot Learning.", + "author": "T\u00f6rnberg, P. 2023a.", + "venue": "arXiv:2304.06588.", + "url": null + } + }, + { + "51": { + "title": "How to use LLMs for Text Analysis.", + "author": "T\u00f6rnberg, P. 2023b.", + "venue": "arXiv:2307.13106.", + "url": null + } + }, + { + "52": { + "title": "Want to reduce labeling cost? GPT-3 can help.", + "author": "Wang, S.; Liu, Y.; Xu, Y.; Zhu, C.; and Zeng, M. 2021.", + "venue": "arXiv:2108.13487.", + "url": null + } + }, + { + "53": { + "title": "Transformers: State-of-the-art natural language processing.", + "author": "Wolf, T.; Debut, L.; Sanh, V.; Chaumond, J.; Delangue, C.; Moi, A.; Cistac, P.; Rault, T.; Louf, R.; Funtowicz, M.; and\u00b4 Sam Shleifer, J. D.; von Platen, P.; Ma, C.; Jernite, Y.; Plu, J.; Xu, C.; Scao, T. L.; Gugger, S.; Drame, M.; Lhoest, Q.; and Rush, A. M. 2020.", + "venue": "In Proceedings of EMNLP.", + "url": null + } + }, + { + "54": { + "title": "Large language models can rate news outlet credibility.", + "author": "Yang, K.-C.; and Menczer, F. 2023.", + "venue": "arXiv:2304.00228.", + "url": null + } + }, + { + "55": { + "title": "Assessing the potential of GPT-4 to perpetuate racial and gender biases in health care: a model evaluation study.", + "author": "Zack, T.; Lehman, E.; Suzgun, M.; Rodriguez, J. A.; Celi, L. A.; Gichoya, J.; Jurafsky, D.; Szolovits, P.; Bates, D. W.; Abdulnour, R.-E. E.; Butte, A. J.; and Alsentzer, E. 2024.", + "venue": "The Lancet Digital Health, 6(1): e12\u2013e22.", + "url": null + } + }, + { + "56": { + "title": "Can ChatGPT Reproduce Human-Generated Labels? A Study of Social Computing Tasks.", + "author": "Zhu, Y.; Zhang, P.; Haq, E.-U.; Hui, P.; and Tyson, G. 2023.", + "venue": "arXiv:2304.10145.", + "url": null + } + }, + { + "57": { + "title": "Can Large Language Models Transform Computational Social Science?", + "author": "Ziems, C.; Held, W.; Shaikh, O.; Chen, J.; Zhang, Z.; and Yang, D. 2023.", + "venue": "Working paper.", + "url": null + } + } + ], + "url": "http://arxiv.org/html/2409.09467v2" +} \ No newline at end of file