Probing the Robustness of Trained Metrics for Conversational Dialogue Systems

TL;DR

This paper presents an adversarial approach using Reinforcement Learning to evaluate the robustness of trained metrics for conversational dialogue systems, revealing their vulnerability to simple, flawed response strategies.

Contribution

It introduces a novel adversarial stress-testing method for trained dialogue metrics, exposing their susceptibility to superficial response strategies.

Findings

Trained metrics can be fooled by simple response strategies

Copying conversation context can yield high metric scores

Metrics often assign high scores to flawed or copied responses

Abstract

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.