[Paper Review] Using Natural Language Explanations to Rescale Human Judgments

Manya Wadhwa, Jifan Chen+2 more|arXiv (Cornell University)|May 24, 2023

Topic Modeling4 citations

TL;DR

This paper proposes a method to rescale coarse-grained Likert-scale human judgments into fine-grained 0–100 scores using natural language explanations and large language models (LLMs), guided by a task-specific scoring rubric. The approach improves alignment with expert judgments without reducing inter-annotator agreement, capturing nuanced subjectivity and scale-use differences.

ABSTRACT

The rise of large language models (LLMs) has brought a critical need for high-quality human-labeled data, particularly for processes like human feedback and evaluation. A common practice is to label data via consensus annotation over human judgments. However, annotators' judgments for subjective tasks can differ in many ways: they may reflect different qualitative judgments about an example, and they may be mapped to a labeling scheme in different ways. We show that these nuances can be captured by natural language explanations, and propose a method to rescale ordinal annotations and explanations using LLMs. Specifically, we feed annotators' Likert ratings and corresponding explanations into an LLM and prompt it to produce a numeric score anchored in a scoring rubric. These scores should reflect the annotators' underlying assessments of the example. The rubric can be designed or modified after annotation, and include distinctions that may not have been known when the original error taxonomy was devised. We explore our technique in the context of rating system outputs for a document-grounded question answering task, where LLMs achieve near-human performance. Our method rescales the raw judgments without impacting agreement and brings the scores closer to human judgments grounded in the same scoring rubric.

Motivation & Objective

• Address the challenge of inconsistent and coarse human judgments in subjective NLP tasks, especially for evaluating LLM outputs.
• Overcome limitations of traditional label aggregation and consensus methods that fail to capture nuanced subjectivity and scale-use differences.
• Leverage natural language explanations from annotators to enhance the expressiveness of numeric scores beyond simple Likert ratings.
• Develop a post-annotation method that enables fine-grained rescaling of human judgments using LLMs and a customizable scoring rubric.
• Improve the quality and fidelity of human feedback for training and evaluating large language models in high-stakes applications.

Proposed method

• Collect human judgments using a Likert scale (e.g., 1–5) and corresponding natural language explanations for each judgment during annotation.
• Feed each annotator’s Likert rating and explanation into a large language model (GPT-4) via a prompt that anchors the rescaling to a predefined, aspect-based scoring rubric.
• Use the LLM to generate a 0–100 numeric score that reflects the annotator’s underlying assessment, informed by both the label and the explanation.
• Design the scoring rubric post-annotation based on analysis of explanations and judgments, enabling distinctions not originally encoded in the original labeling scheme.
• Ensure the rescaling process is consistent and interpretable by grounding the LLM’s output in a structured rubric that defines criteria for each score point.
• Evaluate the rescaled scores against reference human judgments to assess alignment and consistency with expert-level rescaling.

Figure 1: Overview of our method. By feeding explanations that annotators write into an LLM, we can rescale their coarse-grained judgment to a 100-point scale.

Experimental results

Research questions

• RQ1 Can LLM-based rescaling of human judgments, informed by natural language explanations, produce scores that align closely with expert-annotated 0–100 ratings?
• RQ2 Does the proposed method preserve inter-annotator agreement while enabling finer-grained distinctions in judgments?
• RQ3 How effective is the use of a post-annotation scoring rubric in improving the fidelity of rescaled scores compared to baseline rescaling without rubric guidance?
• RQ4 To what extent can natural language explanations capture and convey nuanced subjectivity that is lost in coarse Likert scales?
• RQ5 Can this method be generalized to other subjective NLP evaluation tasks beyond document-grounded question answering?

Key findings

• The proposed explanation-based rescaling (EBR) method produces scores that are significantly closer to expert-annotated 0–100 ratings than baseline rescaling methods.
• Inter-annotator agreement, measured by Krippendorff’s alpha, remains stable after rescaling, indicating that the method does not distort existing consensus.
• The use of a scoring rubric significantly improves rescaling accuracy compared to rescaling without rubric guidance, demonstrating the value of structured task-specific criteria.
• Annotators’ explanations reveal distinct reasoning patterns that are captured and reflected in the rescaled scores, even when they chose the same Likert label.
• The method successfully teases apart differences in scale use and subjectivity, enabling more faithful representation of annotator intent.
• The approach is robust across diverse annotators and maintains interpretability, making it suitable for use in RLHF and other feedback-driven LLM alignment pipelines.

Figure 2: Example from Inquisitive-Broad where annotators were asked to label a question-answer pair on a Likert scale and give explanations for their judgments. Our proposed explanation-based rescaling (EBR) method maps these judgments using an aspect specific scoring rubric to scores that accurate