[Paper Review] Adversarial Evaluation of Dialogue Models

Anjuli Kannan, Oriol Vinyals|arXiv (Cornell University)|Jan 27, 2017

Topic ModelingComputer Science11 references66 citations

TL;DR

The paper studies using a discriminator to evaluate dialogue models by distinguishing machine-generated responses from human ones; the discriminator achieved about 62.5% accuracy and highlighted known weaknesses, but practical adoption remains uncertain.

ABSTRACT

The recent application of RNN encoder-decoder models has resulted in substantial progress in fully data-driven dialogue systems, but evaluation remains a challenge. An adversarial loss could be a way to directly evaluate the extent to which generated dialogue responses sound like they came from a human. This could reduce the need for human evaluation, while more directly evaluating on a generative task. In this work, we investigate this idea by training an RNN to discriminate a dialogue model's samples from human-generated samples. Although we find some evidence this setup could be viable, we also note that many issues remain in its practical application. We discuss both aspects and conclude that future work is warranted.

Motivation & Objective

• Motivate improved evaluation for data-driven dialogue systems beyond perplexity and BLEU.
• Investigate whether an adversarial setup can serve as an automatic proxy for human evaluation.
• Examine what a trained discriminator reveals about strengths and weaknesses of a production-grade dialogue model.

Proposed method

• Generator: a sequence-to-sequence RNN encoder-decoder trained to maximize P(r|o) over observed (o,r) pairs.
• Discriminator: an RNN with an encoder and a binary classifier trained to distinguish (o,r) pairs where r is human versus generated by the generator.
• Training objective for generator: maximize sum over data of log P(r1,...,rm|o1,...,on).
• Training objective for discriminator: maximize sum over data of log P(y|o1,...,on,r1,...,rm) where y indicates human (1) or generator (0).
• Experimental setup mirrors production Smart Reply data; discriminator trained on held-out pairs with half labeled human and half with generator-sampled responses.

Experimental results

Research questions

• RQ1 Can a discriminator effectively distinguish between human and machine-generated dialogue responses in a realistic setting?
• RQ2 What weaknesses or biases of the dialogue model does the discriminator reveal?
• RQ3 Is discriminator-based evaluation aligned with human judgments or improved dialogue quality?
• RQ4 What are the practical challenges in using discriminators for automatic dialogue evaluation?

Key findings

• Discriminator accuracy in distinguishing generator vs human responses is 62.5%.
• Discriminator's decisions reflect known weaknesses: length distribution bias and reliance on simplistic replies like 'Thank you'.
• Discriminator prefers less common language and longer responses, highlighting diversity and length signals as salient features.
• Ranking by discriminator score differs from generator log-likelihood, with weak correlation (Spearman ≈ -0.02) when comparing same-length responses.
• Discriminator performance uncovers weaknesses observed by humans but does not prove a reliable proxy for human evaluation or overall quality.
• The study suggests further investigation and cautions about relying on discriminators as standalone evaluators.