TL;DR
This paper explores using GPT-4 to automatically evaluate the discourse coherence of written essays, a notoriously difficult aspect of writing to assess computationally. They found GPT-4 significantly outperforms traditional NLP methods and achieves strong agreement with human raters, offering a promising avenue for more transparent and efficient automated writing evaluation.
What Problem Does This Paper Address?
Automated Writing Evaluation (AWE) systems have made strides in assessing various aspects of writing quality. However, evaluating "discourse coherence" – how well ideas flow together and create a unified meaning – remains a significant challenge. Traditional AWE systems often rely on surface-level features (like transition words) or complex, less interpretable models. The advent of powerful Large Language Models (LLMs) like GPT-4 presents an opportunity to tackle this challenge, as they can potentially understand and generate text in a way that mimics human judgment. This paper investigates whether GPT-4 can be trained to provide reliable coherence ratings and, crucially, generate human-understandable rationales for those ratings, thereby increasing transparency in AWE.
Methods
The core of this research involves prompting GPT-4 to act as an automated essay scorer for discourse coherence. They developed a detailed prompt that included:
- A description of the task.
- The specific rubric used for scoring.
- Guidelines for terminology and style.
- Examples of previous ratings and rationales (few-shot learning).
- The original prompt given to the test-taker.
- The test-taker's written response.
They experimented with three conditions for GPT-4:
- Rating-first: GPT-4 provides the rating, then the rationale.
- Rationale-first: GPT-4 provides the rationale, then the rating.
- Rating-only: GPT-4 provides only the rating.
As a baseline for comparison, they used traditional NLP coherence metrics based on the Coh-Metrix tool, including:
- Sentence-pair overlap (binary and proportional).
- Coreference resolution.
- Latent Semantic Analysis (LSA) similarity.
These baseline features were used to train a linear regression model to predict human coherence ratings.
Datasets
The study utilized the DET coherence (DET-Coh) dataset, which comprises 500 written responses from the operational Duolingo English Test (DET). This dataset was specifically constructed to ensure an equal distribution of:
- Male and female test-takers.
- Seven major first-language groups.
- Proficiency levels aligned with the Common European Framework of Reference (CEFR).
The dataset was split into:
- 80 responses for interrater agreement checks among human raters.
- 420 responses for the main scoring by human raters.
- 12 responses (two per scale point) for qualitative analysis and few-shot prompting examples.
Key Insights and Ideas
- GPT-4 as a Coherence Judge: The study demonstrates that GPT-4 can effectively assess discourse coherence, achieving agreement levels comparable to human experts. This suggests LLMs are capable of understanding nuanced aspects of writing beyond simple grammar or vocabulary.
- Rationales Enhance, But Don't Dramatically Alter, Performance: While GPT-4's performance was slightly better when it generated rationales (regardless of order), the difference between rating-first, rationale-first, and rating-only conditions was not statistically significant. This implies that the core rating capability is robust, and the addition of rationales primarily adds interpretability.
- Human-like Central Tendency: GPT-4 exhibited a tendency to rate essays closer to the middle of the scale, mirroring a phenomenon observed in human raters (central tendency effect). This might be an artifact of GPT-4 learning from patterns in its vast training data.
- Potential Oversensitivity to Spelling: In cases where GPT-4's ratings differed from human ratings, the model frequently mentioned "spelling" in its rationales. This suggests GPT-4 might be overvaluing spelling errors, potentially due to its token-based representation, which could hinder its ability to discern intended words compared to human readers.
- Bridging the Interpretability Gap: The generated rationales, even if not a perfect reflection of GPT-4's "thought process," offer a significant improvement in transparency for AWE systems. This can help stakeholders understand why a score was given, fostering greater trust and utility.
Key Results
- GPT-4 vs. Baseline: All GPT-4 conditions significantly outperformed the baseline NLP model. For example, the Spearman's rho correlation with human ratings was 0.82 for GPT-4 (in rationale conditions) compared to 0.47 for the baseline.
- Human-GPT-4 Agreement:
- Exact Agreement: GPT-4 achieved 0.51-0.56, significantly higher than the baseline's 0.36.
- Adjacent Agreement: GPT-4 reached 0.95-0.97, far exceeding the baseline's 0.82.
- Quadratic Weighted Kappa (QWK): GPT-4 achieved 0.78-0.82, a substantial improvement over the baseline's 0.39.
- Rationale Impact: The GPT-4 conditions that produced rationales (rating-first and rationale-first) showed marginally, but not significantly, better agreement with human ratings than the rating-only condition.
- Rating Variance: GPT-4's ratings showed less variance than human ratings, with a tendency to produce fewer extreme scores (1s and 6s).
Limitations and Future Work
- Reproducibility and Bias: The authors caution that GPT-4 ratings warrant careful use due to potential issues with reproducibility and inherent biases.
- Limited Insight into "Thought Process": While rationales are valuable for transparency, they may not accurately reflect the underlying decision-making process of GPT-4. They could be "rationalizations" rather than true explanations.
- Spelling Sensitivity: Further research is needed to understand and mitigate GPT-4's potential oversensitivity to spelling errors.
- Prompt Optimization: Future work should focus on further optimizing prompts to elicit even more accurate and interpretable ratings and rationales.
- Broader Applications: The study suggests that LLMs like GPT-4 could be used for more than just scoring, potentially for generating personalized feedback or even creating improved versions of student responses.
- Responsible AI in Education: The research aligns with the broader push for responsible AI in education, emphasizing the need for transparency and fairness in AI-driven assessment tools.
Summary updated