To What Extent Are Large Language Models Capable of Generating Substantial Reflections for Motivational Interviewing Counseling Chatbots? A Human Evaluation

Erkan Basar, Iris Hendrickx, Emiel Krahmer, Gert-Jan Bruijn, Tibor BosseACL 2024

TL;DR

This paper investigates how well large language models (LLMs) like GPT-4 can generate "reflections" for motivational interviewing (MI) chatbots, a crucial technique for helping people change behavior. They found that GPT-4 can produce reflections comparable in quality to human-authored ones, suggesting LLMs have potential to enhance these chatbots, but caution is advised due to occasional inappropriate outputs.

What Problem Does This Paper Address?

Motivational Interviewing (MI) is a therapeutic style that relies heavily on reflective listening to help individuals explore their motivations for behavior change. Creating MI chatbots typically involves pre-written reflections, which can be inflexible and lead to generic responses, hindering empathy and personalization. Automating reflection generation with LLMs could significantly improve these chatbots, but it's critical to ensure the generated reflections are appropriate, specific, natural, and engaging, especially when dealing with sensitive health topics. This paper aims to bridge this gap by evaluating the quality of LLM-generated reflections against human-authored ones in the context of health counseling.

Methods

The researchers conducted a two-step human evaluation.

  1. Independent Evaluation: Human evaluators assessed LLM-generated and human-authored reflections on four key criteria:
  2. Appropriateness: Whether the reflection is ethically and morally suitable for the conversation.
  3. Specificity: Whether the reflection contains information directly from the user's response.
  4. Naturalness: Whether the reflection sounds like it could have been uttered by a person.
  5. Engagement: Whether the reflection encourages further conversation and user interest. A 7-point Likert scale was used for these ratings.

  6. Ranking Evaluation: Evaluators used the RankME method to compare the overall quality of generated reflections against human-authored ones within the same conversation context. This involved assigning a relative score, which was then used to calculate a mean rating value (μ) for each model using the TrueSkill system.

The LLMs evaluated were GPT-4, BLOOM, and FLAN-T5, using zero-shot prompting.

Datasets

The study utilized a collection of human-chatbot conversations in English from two prior studies:

  • Conversations from a smoking cessation chatbot.
  • Conversations from a sexual health chatbot.

These conversations were collected from 175 university students. For the evaluation, 188 conversation contexts were selected where the chatbots were designed to elicit a reflection. No new benchmark datasets were introduced; existing conversational data was repurposed.

Key Insights and Ideas

  • GPT-4 Shines: GPT-4 demonstrated a remarkable ability to generate reflections that were rated highly across all evaluation criteria, often surpassing human-authored reflections. This suggests it's a strong contender for enhancing MI chatbots.
  • LLMs Can Enhance MI: The study confirms that LLMs have the potential to significantly improve the specificity and flexibility of reflections in MI chatbots, moving beyond rigid, pre-written responses.
  • Human-like Quality is Achievable: The findings indicate that LLMs can produce reflections that are not only contextually relevant but also sound natural and engaging, making it difficult to distinguish them from human-generated text.
  • Caution is Crucial: Despite the promising results, the paper strongly emphasizes the need for caution. LLMs can still produce inappropriate or even offensive remarks, especially when encountering unforeseen user responses. A thorough manual review of generated reflections is essential before deployment.
  • Open-Source Models Show Promise, But Lag: While BLOOM, an open-science model, performed comparably to human-authored reflections in independent evaluations, it was ranked lower in the overall quality assessment. FLAN-T5 generally performed the weakest.

Key Results

  • Independent Evaluation: GPT-4 reflections were rated significantly higher than human-authored ones across all four criteria (appropriateness, specificity, naturalness, and engagement) by a 7-point scale. Mean differences ranged from 0.52 to 0.84. BLOOM's reflections showed a similar pattern to human-authored ones, while FLAN-T5 was rated significantly less engaging.
  • Ranking Evaluation: GPT-4 generated reflections achieved the highest overall quality score (μ = 30.46), closely followed by human-authored reflections (μ = 28.05). BLOOM (μ = 25.43) and FLAN-T5 (μ = 20.77) were ranked lower. All pairwise differences in ranking were statistically significant.
  • Qualitative Examples: The paper includes examples where LLMs successfully generated appropriate reflections (Table 5), but also instances where they failed, producing inappropriate responses (Table 3) or struggling with unexpected user input (Table 4).

Limitations and Future Work

  • Context Length: The study evaluated reflections based on a limited context of 5 preceding turns. Future work could explore longer contexts or conversation memory to assess how LLMs handle more complex interactions and maintain engagement over time.
  • Participant Diversity: The study used non-experts for evaluation and did not include participants who were actively seeking to quit smoking or address sexual health issues. Future research could involve target users to gain more direct feedback.
  • Post-processing: BLOOM and FLAN-T5 required additional automated post-processing to remove repetitive or near-duplicate outputs. Future work aims to incorporate a wider range of LLMs and compare them more directly.
  • Manual Review Necessity: The authors strongly recommend a thorough manual review of LLM-generated reflections before their use in sensitive counseling contexts, as there's no guarantee of consistent appropriateness.
  • Broader LLM Evaluation: The study focused on a few specific LLMs. Future research could expand to include a wider array of open-source and proprietary models.

Summary updated

BibTeX
@inproceedings{basar2024what,
  author = {Erkan Basar and Iris Hendrickx and Emiel Krahmer and Gert-Jan Bruijn and Tibor Bosse},
  title = {{To What Extent Are Large Language Models Capable of Generating Substantial Reflections for Motivational Interviewing Counseling Chatbots? A Human Evaluation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure