Source-linked AI summary
Adversarial Evaluation of Dialogue Models
Anjuli Kannan, Oriol Vinyals
TL;DR
Dialogue evaluation remains challenging because common automatic measures miss task-specific response quality and human evaluation does not scale well. The paper trains an RNN discriminator to distinguish generated from human replies, finding that it can do so over 60% of the time. The authors also identify unresolved practical issues and call for further investigation.
Problem
Dialogue evaluation remains challenging: perplexity does not measure task performance, BLEU poorly fits varied valid replies, and human evaluation does not scale well.
Source: §1 · p. 1Method
The authors keep a production-scale dialogue generator fixed and train an RNN discriminator to distinguish its responses from human responses to the same messages.
Source: §1 · p. 1Results
The discriminator distinguishes model output from human output over 60% of the time and appears to detect weaknesses involving response length and familiar, simplistic replies.
Source: §1 · p. 1Takeaways & Limitations
Abstract
from arXiv · showhide
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.
1 Introduction
Dialogue evaluation remains difficult because common automatic measures miss task-appropriate replies and human review does not scale well. The paper explores an RNN discriminator as an adversarial evaluation method, finding it distinguishes model from human responses while leaving its practical value unresolved.
1 Introduction
- RNN encoder-decoder dialogue models had improved over statistical machine translation and rule-based chatbots, while evaluation remained a challenge.Source: §1 · p. 1
- Perplexity does not measure task performance, BLEU poorly fits dialogue’s varied valid replies, and human evaluation is costly and sometimes unavailable.The paper cites Smart Reply [3] as a setting where data cannot be viewed by humans.Source: §1 · p. 1
- The proposed evaluation trains a discriminator to distinguish a fixed production-scale dialogue model’s responses from human responses to the same messages.The aim is to assess whether this adversarial setup is viable for evaluation.Source: §1 · p. 1
- The discriminator distinguishes model output from human output over 60% of the time, but lower discriminator accuracy is not shown to imply better human evaluations.It also appears to detect weaknesses humans have observed: response-length distribution and reliance on familiar, simplistic replies.Source: §1 · p. 1
1.1 Related work
- Unlike prior adversarial evaluation [1], this work examines discriminator properties on a real-world task; its GAN-inspired setup addresses sequential, conditional dialogue and keeps the models separate.The paper concludes that further investigation is warranted.Source: §1.1 · p. 2, §1 · p. 2
2 Model
The model uses separate RNN-based generator and discriminator networks rather than jointly training them toward one objective. The generator predicts responses from messages, while the discriminator predicts whether a response is human or generated.
2 Model
- The generator and discriminator are separate models, not jointly trained to a single objective.Source: §2 · p. 2
- The generator is an RNN encoder-decoder trained to maximize the log probability of observed responses given their original messages.Its input is a message token sequence and its output is a response token sequence.Source: §2 · p. 2
- The discriminator uses an RNN encoder and binary classifier to predict whether a response paired with its message came from training data.Its objective is to maximize log P(y|o1, ..., on, r1, ..., rm).Source: §2 · p. 2
3 Experiments
The experiments train a discriminator on held-out email replies paired with either human or generator responses. It distinguishes the sources 62.5% of the time and shows preferences associated with response length and diversity, though those preferences do not establish response quality.
3.1 Data and training
- The discriminator is trained on held-out email pairs, half unchanged as human examples and half paired with generator-sampled responses.The generator is trained on the same email data and in the same manner as the Smart Reply production model [3].Source: §3.1 · p. 2
3.2 Discriminator performance
- 62.5% of the time, the discriminator distinguishes generator samples from human samples conditional on the original message.Source: §3.2 · p. 2
- Figure 1 shows discriminator score versus length and recall versus precision, including the precision-recall curve for the discriminator.Source: Figure 1 · p. 3, §3.2 · p. 3
3.3 Comparison with perplexity
- Discriminator preferences correlate strongly with response length; among equal-length responses, rankings differ from generator likelihood (average Spearman’s correlation -0.02) and favor less-common replies.The discriminator favors rarer language, including examples such as “Ya” over “Yes,” and omits “Thank you” from its top responses; this does not establish that rarer replies are better.Source: §3.3 · p. 3, Table 1 · p. 3
4 Discussion
The discussion presents per-system discriminators as the simplest proposed alternative to costly human evaluation and learned critics. Preliminary results are promising, but the mismatch between training on samples and using maximally likely responses in practice leaves application questions unresolved.
4 Discussion
- The discussion contrasts costly human evaluation and potentially overfit learned critics with per-system discriminators that could score new systems using varied critics.It presents the discriminator approach as the simplest of these three evaluation paths.Source: §4 · p. 4
- The critic identifies sequence length and diversity weaknesses in a production-quality dialogue system and distinguishes its real responses from generated ones.The discussion describes these as preliminary findings.Source: §4 · p. 4
- Discriminators trained on samples face a mismatch with practice, which uses maximally likely responses or approximations, and further work is needed.The passage notes that discriminators have a harder time with sampling than beam search, despite human observations that some search is useful for high-quality responses.Source: §4 · p. 4