Ask the experts: sourcing a high-quality nutrition counseling dataset through Human-AI collaboration

Simone Balloccu, Ehud Reiter, Karen Li, Rafael Sargsyan, Vivek Kumar, Diego Reforgiato Recupero, Daniele Riboni, Ondrej DusekEMNLP 2024 · DOI

Introduction

As Large Language Models (LLMs) become more integrated into daily life, many individuals are turning to them for sensitive health-related guidance, including nutrition counseling. Nutrition counseling is not merely about calculating calories; it requires "soft skills" such as empathy, reframing, and supportive dialogue to help individuals navigate complex dietary struggles. However, current LLMs often lack the domain-specific training necessary to provide safe and effective psychological support in this arena.

To address this gap, researchers have introduced HAI-Coaching, the first public dataset specifically designed for nutrition counseling. This dataset bridges the divide between fluent AI generation and expert-level health coaching by leveraging a Human-AI (HAI) collaboration framework.

The HAI-Coaching Dataset

The HAI-Coaching dataset consists of approximately 2,400 real-world dietary struggles crowdsourced from over 800 individuals. These struggles cover 12 distinct topics, ranging from emotional eating and craving habits to situational challenges and mental health-related dietary issues.

To provide comprehensive support, the researchers defined four categories of "supportive text":

  1. Reflection: Summarizing the user's struggle to show understanding.
  2. Comfort: Normalizing the experience so the user does not feel alone or guilty.
  3. Reframing: Offering a positive or constructive perspective on the situation.
  4. Suggestion: Providing practical, actionable advice.
An example prompt template used to generate 'Comfort' responses, designed to normalize the user's struggle.

Methodology: Human-AI Collaboration

The creation of HAI-Coaching relied on a sophisticated iterative loop between AI generation and expert oversight. A team of 13 nutrition experts worked to refine prompts for ChatGPT (GPT-3.5), ensuring that the generated responses adhered to professional counseling standards.

The Human-AI (HAI) collaboration framework used to develop the HAI-Coaching dataset.

The process involved several key stages:

  • Requirement Analysis: Defining the types of support needed with expert input.
  • Struggle Collection: Gathering authentic user experiences from platforms like Prolific.
  • Prompt Engineering: Iteratively updating prompts based on expert feedback to improve response quality.
  • Mass Generation: Creating nearly 97,000 candidate responses across the four supportive categories.
  • Safety Annotation: Experts manually labeled responses as safe or unsafe, identifying risks that automated metrics often overlook.

Key Findings: The "Safe but Useless" Trap

One of the most significant insights from the study is the prevalence of "safe but useless" content. While 85% of ChatGPT's responses were technically "safe" (i.e., not directly harmful), experts found many to be generic or dismissive of the underlying psychological roots of a struggle. For instance, a model might suggest "setting a reminder" for someone struggling with deep-seated emotional eating—a suggestion that is safe but ultimately ineffective.

Furthermore, the study highlighted the risk of AI reinforcing "toxic diet culture." In some instances, the model suggested using food as a reward for diet adherence, a practice experts identify as a potential precursor to eating disorders. This underscores the danger of deploying LLMs in health contexts without rigorous expert-led safety guardrails.

Benchmarking and Performance

The researchers evaluated several open-source models—including Mistral 7B, Llama 3 8B, and Phi 3 mini—on tasks like struggle classification and safety detection.

  • Classification: Fine-tuned smaller models like RoBERTa performed surprisingly well, rivaling much larger LLMs in categorizing struggles.
  • Safety Detection: This remained the most challenging task, with models achieving only moderate accuracy (~0.66–0.69). The rarity of "unsafe" examples (15%) makes it difficult for models to learn the subtle nuances of psychological harm.
  • Generation: While Phi 3 mini showed strong performance on automatic metrics like BLEURT, manual inspection revealed persistent issues with repetitiveness and a lack of genuine empathy.

Conclusion and Future Directions

The HAI-Coaching dataset represents a major step toward safer AI-driven health coaching. However, the study concludes that current LLMs are not yet ready for unsupervised nutrition counseling. Future work must focus on expanding the dataset to include more diverse cultural perspectives and specific medical conditions (like diabetes or pregnancy). Ultimately, the goal is to move beyond "safe" text toward AI that is truly empathetic, useful, and medically sound.

Summary updated

BibTeX
@inproceedings{balloccu2024ask,
  author = {Simone Balloccu and Ehud Reiter and Karen Li and Rafael Sargsyan and Vivek Kumar and Diego Reforgiato Recupero and Daniele Riboni and Ondrej Dusek},
  title = {{Ask the experts: sourcing a high-quality nutrition counseling dataset through Human-AI collaboration}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  doi = {10.18653/v1/2024.findings-emnlp.674},
}
Figure