Comparing human text classification performance and explainability with large language and machine learning models...

Area of Science:

• Artificial Intelligence
• Human-Computer Interaction
• Natural Language Processing

Background:

• Comparing artificial intelligence (AI) models with human reasoning is crucial for AI development.
• Text classification tasks, particularly with complex datasets, pose challenges for both humans and AI.

Purpose of the Study:

• To empirically compare the text classification performance and explainability of humans, a traditional machine learning (ML) model, and a large language model (LLM).
• To investigate the alignment between human and AI reasoning in a real-world classification task.

Main Methods:

• A user study with 51 participants classifying 204 injury narratives into 6 cause-of-injury codes, with eye-tracking data recorded.
• Trained ML model using 120,000 pre-labeled narratives; LLM and humans received no specialized training.
• Explainability compared using top classification decision words identified via eye-tracking (humans), LIME (ML), and prompts (LLM).

Main Results:

• The ML model demonstrated superior classification performance compared to zero-shot LLM and non-expert humans, particularly for complex and difficult-to-categorize narratives.
• Both ML and LLM showed higher agreement with human reasoning in their top-3 predictive words than with later predictive words.
• Explainability analysis revealed varying degrees of alignment in decision-making processes.

Conclusions:

• Traditional ML models can still outperform LLMs in specific, domain-specific text classification tasks, especially with sufficient training data.
• While LLMs show promise, human-level performance and explainability in complex classification tasks require further research and development.
• Understanding the alignment of reasoning between humans and AI is key to building trustworthy and effective AI systems.