A protocol for collecting speech data with varying degrees of trust

Lara Gauder, Agustín Gravano, Luciana Ferrer, Pablo Riera, Silvina BrussinoSMM 2019 · DOI

TL;DR

This paper introduces a novel experimental protocol designed to collect speech data from users interacting with a virtual assistant under varying degrees of induced trust. The goal is to create a dataset that can be used to automatically predict a user's trust level from their speech, a crucial aspect for improving human-AI interaction.

What Problem Does This Paper Address?

As conversational agents become more sophisticated, understanding and responding to a user's mental state is paramount. One critical aspect of this mental state is the user's trust in the agent's capabilities. If a system can detect a user's distrust, it could potentially adapt its behavior to regain that trust. However, there's a significant lack of speech datasets annotated with varying degrees of trust, making it difficult to develop automatic trust detection systems. Existing methods for annotating trust from speech have proven unreliable, suggesting a need for a controlled environment to generate such data.

Methods

The core of this paper is a carefully designed experimental protocol. Participants interact with a virtual assistant (VA) through a text-based chat interface, answering factual questions. The key innovation lies in how trust is manipulated:

  • Initial Bias (Anchoring Effect): Before the interaction, participants are told that the VA has been previously rated by other users as either very good (high-score condition) or very bad (low-score condition). This leverages the anchoring bias, influencing their initial perception of the VA.
  • Reinforced Performance: The VA's performance is then tailored to match the initial bias. In the high-score condition, the VA answers all questions correctly. In the low-score condition, it makes some deliberate errors on difficult questions, reinforcing the initial negative impression without completely alienating the user.
  • Task Structure: Participants are asked to:
  • Interact with the VA to find the answer.
  • Transcribe the VA's response.
  • Rate their confidence in the VA's answer (5-level Likert scale).
  • Provide their own answer.
  • Evaluation Surveys: Participants complete brief surveys at specific points to measure their perceived trust and confidence in the VA.

Datasets

The paper describes a preliminary experiment using a text-only version of the protocol in Argentine Spanish. The results are based on data collected from 15 volunteers. The authors plan to make the resulting speech dataset publicly available once it's collected. No existing datasets are used for training or evaluation in this preliminary phase; the focus is on generating a new, controlled dataset.

Key Insights and Ideas

  • The Power of Anchoring: The initial information about the VA's ratings significantly influences participants' subsequent trust and confidence, demonstrating the effectiveness of the anchoring bias in experimental design.
  • Trust is Multifaceted: The study uses multiple measures (overall survey ratings, confidence in specific answers, and re-use of VA responses) to capture different facets of trust and reliance.
  • Controlled Trust Manipulation is Feasible: The protocol successfully induces different levels of trust, providing a controlled environment for collecting speech data that reflects these varying states.
  • Speech as a Trust Indicator: The hypothesis that speech characteristics (intonation, pitch, speech rate, voice quality) can reveal a user's trust level is the driving force behind this research.

Key Results

The preliminary text-only experiment yielded promising results:

  • Survey Ratings: Participants in the high-score condition reported a significantly higher mean degree of trust (4.42) compared to the low-score condition (3.39).
  • Confidence in Answers: Confidence in the VA's answers for difficult questions was also higher in the high-score condition (mean 4.34) versus the low-score condition (mean 3.71).
  • Response Re-use: Participants in the high-score condition reused the VA's answers for difficult questions 91.7% of the time, compared to 83.3% in the low-score condition, indicating greater reliance on the VA when trust was higher.
  • These differences were statistically significant, supporting the protocol's effectiveness in influencing participants' perceived trust.

Limitations and Future Work

  • Preliminary Nature: The current experiment is text-only and serves as a proof-of-concept. The real value will come from the speech data collection.
  • Individual Differences: Not all participants showed clear differences in trust between conditions, suggesting that individual factors might play a role. Further analysis is needed to understand these variations.
  • Future Work: The immediate next step is to implement the speech-based version of the protocol to collect the target dataset. Once collected, the authors aim to:
  • Investigate whether speakers exteriorize trust through their speech.
  • Develop machine learning models to automatically detect trust from speech.
  • Formalize the pilot study on external annotation of trust to assess its feasibility on the new dataset.

Summary updated

BibTeX
@inproceedings{gauder2019protocol,
  author = {Lara Gauder and Agustín Gravano and Luciana Ferrer and Pablo Riera and Silvina Brussino},
  title = {{A protocol for collecting speech data with varying degrees of trust}},
  booktitle = {SMM},
  year = {2019},
  doi = {10.21437/smm.2019-2},
}
Figure