ElicitationGPT: Text Elicitation Mechanisms via Language Models

The paper introduces a framework for evaluating textual information using large language models, creating proper scoring rules that align with human judgment to incentivize truthful reporting, and demonstrates its effectiveness through empirical evaluation on a peer-grading dataset.

Abstract

Scoring rules evaluate probabilistic forecasts of an unknown state against the realized state and are a fundamental building block in the incentivized elicitation of information and the training of machine learning models. This paper develops mechanisms for scoring elicited text against ground truth text using domain-knowledge-free queries to a large language model (specifically ChatGPT) and empirically evaluates their alignment with human preferences. The empirical evaluation is conducted on peer reviews from a peer-grading dataset and in comparison to manual instructor scores for the peer reviews.

Method

The authors used a methodology that involved constructing proper scoring rules for text and evaluating their alignment with human preferences. This included mapping text into a high-dimensional semantic space and employing domain-knowledge-free queries to a large language model (specifically ChatGPT). The empirical evaluation was conducted on a peer-grading dataset, comparing the scoring rules' alignment with manual instructor scores for the peer reviews. The study also considered the correlation of the scoring rules with students' overall grades to assess their reliability and potential for use in educational settings.

Main Finding

The authors discovered that their textual scoring rules, which were designed to evaluate peer reviews in a high-dimensional semantic space using large language models, showed a high degree of alignment with human evaluators (instructors). Furthermore, they found that these scoring rules were more aligned with students' overall grades than the instructors' scores, suggesting that the automated scoring rules might be more reliable than human grading in some contexts. This indicates potential for the use of such rules in large-scale courses to assist with peer grading and reduce the workload on instructors while maintaining or improving the accuracy of grading.

Conclusion

The conclusion of the paper by Yifan Wu and Jason Hartline is that their proposed framework for constructing proper scoring rules to evaluate textual information using large language models is effective and aligns well with human preferences. The empirical evaluation on a peer-grading dataset demonstrated that the scoring rules not only correlate highly with instructor scores but also show a strong alignment with students' overall grades, potentially making them a reliable tool for assisting in peer grading processes in educational settings.

Keywords

Scoring rules, text elicitation, large language models, ChatGPT, proper scoring rules, peer grading, machine learning, semantic space, domain-knowledge-free queries, empirical evaluation, alignment with human preferences, grading mechanism, incentivized information elicitation, loss functions, calibration, prediction accuracy, human computation, mechanism design, language model oracles, summarization, question answering, properness, multi-dimensional aggregation, know-it-or-not beliefs, filtered average aggregation, robustness, manipulation resistance.