Stakeholder feedback is essential for policymakers to evaluate and develop effective policies, but traditional qualitative analysis methods are often labor-intensive and time-consuming. This study investigates the use of Large Language Models (LLMs) like GPT-4 Turbo (GPT-4) with human expertise to analyze stakeholder interviews regarding K–12 education policy in a U.S. state. The research employed a mixed-methods approach where human experts developed a codebook and iterative prompts for GPT-4 to conduct thematic and sentiment analysis. Results demonstrated that GPT-4’s thematic coding achieved 78% agreement with human coding at detailed levels and 96% alignment for broader themes, exceeding traditional Natural Language Processing methods by over 25%. GPT-4 also produced sentiment analysis results more closely aligned with a human expert’s judgment. Our qualitative comparisons between human and GPT-4 analysis results highlight the complementary roles of human expertise and LLMs in enhancing efficiency, validity, and interpretability of educational policy research.
more »
« less
This content will become publicly available on January 1, 2027
Data Plus Theory Equals Codebook: Leveraging LLMs for Human-AI Code-book Development
Recent research has explored the use of Large Language Models (LLMs) to develop qualitative codebooks, mainly for inductive work with large datasets, where manual review is impractical. Although these efforts show promise, they often neglect the theoretical grounding essential to many types of qualitative analysis. This paper investigates the potential of GPT-4o to support theory-informed codebook development across two educational contexts. In the first study, we employ a three-step approach—drawing on Winne & Hadwin's and Zimmerman's Self-Regulated Learning (SRL) theories, think-aloud data, and human refinement—to evaluate GPT-4o's ability to generate high-quality, theory-aligned codebooks. Results indicate that GPT-4o can effectively leverage its knowledge base to identify SRL constructs reflected in student problem-solving behavior. In the second study, we extend this approach to a STEM game-based learning context guided by Hidi & Renninger's four-phase model of Interest Development. We compare four prompting strategies: no theories provided, theories named, full references given, and full-text theory papers supplied. Human evaluations show that naming the theory without including full references produced the most practical and usable codebook, while supplying full papers to the prompt enhanced theoretical alignment but reduced applicability. These findings suggest that GPT-4o can be a valuable partner in theory-driven qualitative research when grounded in well-established frameworks, but that attention to prompt design is required. Our results show that widely available foundation models—trained on large-scale open web and licensed datasets—can effectively distill established educational theories to support qualitative research and codebook development. The code for our codebook development process and all the employed prompts and codebooks produced by GPT are available for replication purposes at: https://osf.io/g3z4x.
more »
« less
- PAR ID:
- 10691493
- Publisher / Repository:
- Zenodo
- Date Published:
- Journal Name:
- Journal of educational data mining
- ISSN:
- 2157-2100
- Subject(s) / Keyword(s):
- large language models qualitative codebooks interest development self-regulated learning thematic analysis codebook development
- Right(s):
- Creative Commons Attribution 4.0 International
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Language-based text provide valuable insights into people’s lived experiences. While traditional qualitative analysis is used to capture these nuances, new paradigms are needed to scale qualitative research effectively. Artificial intelligence presents an unprecedented opportunity to expand the sale of analysis for obtaining such nuances. The study tests the application of GPT-4—a large language modeling—in qualitative data analysis using an existing set of text data derived from 60 qualitative interviews. Specifically, the study provides a practical guide for social and behavioral researchers, illustrating core elements and key processes, demonstrating its reliability by comparing GPT-generated codes with researchers’ codes, and evaluating its capacity for theory-driven qualitative analysis. The study followed a three-step approach: (1) prompt engineering, (2) reliability assessment by comparison of GPT-generated codes with researchers’ codes, and (3) evaluation of theory-driven thematic analysis on psychological constructs. The study underscores the utility of GPT’s capabilities in coding and analyzing text data with established qualitative methods while highlighting the need for qualitative expertise to guide GPT applications. Recommendations for further exploration are also discussed.more » « less
-
We introduce the GEM (Generative Estimator for Mutual Information), an evaluation metric for assessing language generation by large language models (LLMs), particularly in generating informative judgments, without the need for a gold standard reference. GEM broadens the scenarios where we can benchmark LLM generation performance-from traditional ones, like machine translation and summarization, where gold standard references are readily available, to subjective tasks without clear gold standards, such as academic peer review. GEM uses a generative model to estimate mutual information between candidate and reference responses, without requiring the reference to be a gold standard. In experiments on two human-annotated datasets, GEM demonstrates competitive correlations with human scores compared to the state-of-the-art GPT-4o Examiner, and outperforms all other baselines. Additionally, GEM is more robust against strategic manipulation, such as rephrasing or elongation, which can artificially inflate scores under a GPT-4o Examiner. We also present GRE-bench (Generating Review Evaluation Benchmark) which evaluates LLMs based on how well they can generate high-quality peer reviews for academic research papers. Because GRE-bench is based upon GEM, it inherits its robustness properties. Additionally, GRE-bench circumvents data contamination problems (or data leakage) by using the continuous influx of new open-access research papers and peer reviews each year. We show GRE-bench results of various popular LLMs on their peer review capabilities using the ICLR2023 dataset.more » « less
-
This full research paper evaluates the capabilities of large language models (LLMs) in cybersecurity education through two scenario-based multiple-choice questions concept inventories: the Cybersecurity Concept Inventory (CCI) and the Cybersecurity Curriculum Assessment (CCA). The test models included Llama3, Claude 3.5, DeepSeek, and OpenAI's GPT series (GPT-3.5, GPT-4, GPT-4o, and GPT-o1). We evaluated these LLMs in a zero-shot setting to minimize external influencing factors and compared the LLMs' performance to that of students previously studied. We also applied the LLMs' responses to the derived Item Response Theory (IRT) model to compare their knowledge levels. Quantitative results revealed that GPT-o1 outperformed all other models and average students in correctness and knowledge level, correctly answering 92% of the CCI and 72% of CCA test items. Additionally, the research investigates whether GPT-o1 makes mistakes through expert review of GPTo1's output. Expert review of GPT-o1's reasoning patterns uncovered instances of insightful reasoning but also highlighted ways in which GPT-o1's answers reflect persistent student mistakes, such as biases, overgeneralizations, and logical inconsistencies. This work highlights the great potential of GPT-o1 for introductory cybersecurity education for providing detailed explanations and structured reasoning for novice learners.more » « less
-
Scientific progress depends on the ability of researchers to synthesize the growing body of literature. Can large language models (LLMs) assist scientists in this task? Here we introduce OpenScholar, a specialized retrieval-augmented language model (LM)1 that answers scientific queries by identifying relevant passages from 45 million open-access papers and synthesizing citation-backed responses. To evaluate OpenScholar, we develop ScholarQABench, the first large-scale multi-domain benchmark for literature search, comprising 2,967 expert-written queries and 208 long-form answers across computer science, physics, neuroscience and biomedicine. Despite being a smaller open model, OpenScholar-8B outperforms GPT-4o by 6.1% and PaperQA2 by 5.5% in correctness on a challenging multi-paper synthesis task from the new ScholarQABench. Although GPT-4o hallucinates citations 78–90% of the time, OpenScholar achieves citation accuracy on par with human experts. OpenScholar’s data store, retriever and self-feedback inference loop improve off-the-shelf LMs: for instance, OpenScholar-GPT-4o improves the correctness of GPT-4o by 12%. In human evaluations, experts preferred OpenScholar-8B and OpenScholar-GPT-4o responses over expert-written ones 51% and 70% of the time, respectively, compared with 32% for GPT-4o. We open-source all artefacts, including our code, models, data store, datasets and a public demo.more » « less
An official website of the United States government
