LLMs in Short Answer Scoring: Limitations and Promise of Zero-Shot and Few-Shot Approaches

Imran Chamieh, Torsten Zesch, Klaus GiebermannNAACL 2024

TL;DR

This paper investigates the effectiveness of Large Language Models (LLMs) like GPT and LLaMA for automated short answer scoring using zero-shot and few-shot learning. While LLMs show promise on general knowledge tasks, they struggle with complex reasoning and domain-specific knowledge, indicating they are not yet ready for widespread adoption in this area without fine-tuning.

What Problem Does This Paper Address?

Automated scoring of short answers is crucial for scaling educational assessments and providing timely feedback. Traditional methods often require extensive labeled training data, which is costly and time-consuming to acquire. This paper explores whether LLMs, with their vast pre-training, can overcome this data bottleneck by performing well in zero-shot (no examples) and few-shot (a few examples) settings for short answer scoring. The goal is to understand their limitations and potential for this task, especially when training data is scarce.

Methods

The researchers evaluated two prominent LLMs: OpenAI's GPT family (specifically GPT-3.5 and GPT-4) and Meta AI's LLaMA models (LLaMA-7b, LLaMA-13b, and LLaMA-70b). They compared these LLMs in three settings:

  • Zero-shot: The LLM receives only the prompt and the student's answer, with no prior examples of scored answers.
  • Few-shot: The LLM receives the prompt, the student's answer, and a small number of example answer-score pairs. They experimented with different numbers of shots (1, 3, 5, and 10) and various prompt delimiters (newline, semicolon, space).
  • Fine-tuning: A GPT-3.5 model was fine-tuned on a portion of the training data to adapt it specifically to the scoring task.

Performance was benchmarked against established methods like BERT and Support Vector Machines (SVMs). The primary evaluation metric was Quadratic Weighted Kappa (QWK), which measures agreement between machine and human expert scores.

Datasets

The study utilized three publicly available short-answer scoring datasets:

  • ASAP (Automated Student Assessment Prize): A broad dataset with 10 prompts across various disciplines, scored by two humans on a 0-2 or 0-3 scale.
  • MindReading: Contains responses from children (ages 7-13) to questions from specific tasks, scored on a 0-2 scale. The questions themselves were not available to the LLMs.
  • Powergrading: A dataset focused on U.S. citizenship exam knowledge, with answers scored on a 0-1 scale.

Key Insights and Ideas

  • General Knowledge is Key: LLMs perform significantly better on tasks requiring general knowledge (like the U.S. citizenship test in Powergrading) compared to those demanding complex reasoning or domain-specific expertise (like ASAP).
  • Zero-Shot Struggles: LLMs are not yet mature enough for reliable zero-shot automated scoring across the board. Their performance is highly dataset-dependent.
  • Few-Shot Offers Limited Gains: While few-shot learning can provide a slight improvement for some LLMs (particularly LLaMA models on Powergrading and MindReading), it doesn't consistently yield substantial gains, and in some cases (GPT-3.5), performance can even degrade with more shots.
  • Prompt Engineering Matters (But Not Always Dramatically): Different prompt delimiters showed minimal variance in results, suggesting that while prompt design is important, it might not be the sole factor determining LLM performance in this context.
  • Fine-tuning is Powerful but Costly: Fine-tuning LLMs can bring their performance close to supervised baselines, but it comes with significantly higher computational costs.
  • Overfitting Concerns: The fine-tuned GPT-3.5 model showed signs of overfitting on certain tasks, consistently assigning the same score, which is a red flag for robust automated scoring.

Key Results

  • Zero-Shot: GPT-4 showed promising results on Powergrading (QWK up to .86), but all models performed poorly on ASAP (QWK as low as .029 for GPT-3.5).
  • Few-Shot: LLaMA models showed a slight improvement with up to 6 shots on Powergrading. GPT-3.5's performance decreased with more shots on Powergrading.
  • Fine-tuning: The fine-tuned GPT-3.5 model's performance was comparable to supervised results but was computationally more expensive. Interestingly, BERT scored slightly higher QWK than fine-tuned GPT-3.5 across datasets, suggesting potential overfitting in the LLM.
  • Baselines: Supervised methods and BERT generally outperformed zero-shot and few-shot LLM approaches on most datasets, especially ASAP.

Limitations and Future Work

Limitations: * The study only tested commercial LLMs, and the authors hypothesize that open-source LLMs might perform even worse. * The paper acknowledges that LLMs are still not ready for zero-shot or few-shot use in short answer scoring without significant improvements. * LLMs can be "black boxes," making it difficult to interpret why they assign a particular score. * Ethical concerns regarding potential biases in LLM training data and student data privacy when using hosted LLMs are highlighted.

Future Work: * The authors plan to test a wider range of LLMs to gain further insights into their capabilities. * Further research is needed to improve LLMs' ability to handle nuanced and specialized content, as well as complex reasoning.

Summary updated

BibTeX
@inproceedings{chamieh2024llms,
  author = {Imran Chamieh and Torsten Zesch and Klaus Giebermann},
  title = {{LLMs in Short Answer Scoring: Limitations and Promise of Zero-Shot and Few-Shot Approaches}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure