Probing the Robustness of Trained Metrics for Conversational Dialogue Systems
This work addresses the reliability of evaluation metrics for conversational AI, highlighting a critical vulnerability that could mislead system development.
The paper tackles the problem of evaluating conversational dialogue systems by stress-testing trained metrics using an adversarial reinforcement learning method, finding that these metrics give high scores to simple, flawed strategies like copying conversation context, which can outperform human-written responses.
This paper introduces an adversarial method to stress-test trained metrics to evaluate conversational dialogue systems. The method leverages Reinforcement Learning to find response strategies that elicit optimal scores from the trained metrics. We apply our method to test recently proposed trained metrics. We find that they all are susceptible to giving high scores to responses generated by relatively simple and obviously flawed strategies that our method converges on. For instance, simply copying parts of the conversation context to form a response yields competitive scores or even outperforms responses written by humans.