[Paper Review] Beyond Turing: Intelligent Agents Centered on the User

Maxine Eskénazi, Shikib Mehri+2 more|arXiv (Cornell University)|Jan 20, 2019

Speech and dialogue systems36 references10 citations

TL;DR

This paper proposes a user-centric paradigm for intelligent agents, shifting focus from agent performance to user impact. It introduces User-Feedback-Driven Assessment (NUF) to evaluate system responses based on subsequent user reactions, demonstrating that models predicting user satisfaction from user utterances alone achieve 65.8% accuracy and 0.67 MAE, proving user feedback is the most reliable indicator of system quality.

ABSTRACT

Most research on intelligent agents centers on the agent and not on the user. We look at the origins of agent-centric research for slot-filling, gaming and chatbot agents. We then argue that it is important to concentrate more on the user. After reviewing relevant literature, some approaches for creating and assessing user-centric systems are proposed.

Motivation & Objective

• Address the imbalance in dialog research that prioritizes agent capabilities over user experience.
• Challenge the long-standing focus on human-like deception (e.g., Turing test) as a benchmark for success.
• Develop assessment methods that prioritize user satisfaction and real-time feedback over traditional metrics.
• Demonstrate that user feedback following a system response is the most reliable signal for evaluating system quality.
• Provide a framework for designing agents that serve users as partners, not just as functional tools or conversational imitators.

Proposed method

• Introduce a user-feedback-driven assessment framework where system response quality is judged by the user's next utterance.
• Define a three-way quality scale (S_sys < S_usr, S_sys = S_usr, S_sys > S_usr) to compare system response quality against user reaction.
• Use a subset of DSTC-1 data with 1250 annotated c-x-u (context-system-user) triples, achieving Fleiss Kappa of 0.515 for user feedback.
• Train both classification and regression models using TF-IDF weighted n-gram features (n=1,2) on concatenated context, system, and user utterances.
• Evaluate model performance using accuracy and mean absolute error (MAE), with regression preferred due to ordinal nature of user feedback.
• Explore a universal user satisfaction model that uses only the user's next utterance as input, showing high performance with 60.3% accuracy and 0.81 MAE.

Experimental results

Research questions

• RQ1 How can we assess intelligent agent performance in a way that prioritizes user experience over agent-centric metrics?
• RQ2 To what extent does the user's next utterance serve as a reliable indicator of the quality of the system's previous response?
• RQ3 Can a model trained solely on user utterances predict user satisfaction with high accuracy, reducing reliance on costly full dialog annotations?
• RQ4 What are the limitations of traditional evaluation methods that assess system responses in isolation, without considering user feedback?
• RQ5 How can we design agent systems that function as true partners by dynamically adapting to user needs based on real-time feedback?

Key findings

• The model that uses only the user's next utterance (u) as input achieves 60.3% accuracy and 0.81 MAE in predicting user satisfaction, indicating strong predictive power.
• The full model using context, system response, and user utterance (c-x-u) achieves 65.8% accuracy and 0.67 MAE, outperforming models using only context and system response.
• The difference between S_sys < S_usr (18.6%) and S_sys > S_usr (35.0%) highlights that isolated assessment overestimates system quality by nearly double.
• Regression models outperform classification for this task due to the ordinal nature of user feedback, reducing confusion between adjacent quality levels.
• The high Fleiss Kappa of 0.515 for user feedback annotations confirms inter-annotator agreement is substantial, validating the reliability of the NUF metric.
• Using the user's next utterance as a signal for system quality leads to more accurate and user-centered evaluation than traditional context-only or response-only assessments.