Explaining Dialogue Evaluation Metrics using Adversarial Behavioral Analysis

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5871 - 5883 July 10-15, 2022 ©2022 Association for Computational Linguistics Explaining Dialogue Evaluation Metrics using Adversarial Behavioral Analysis Baber Khalid Rutgers University 110 Frelinghuysen Rd Piscataway, NJ 08854 baber.khalid@rutgers.edu Sungjin Lee Amazon Alexa 300 Pine Street Seattle, WA 98103 sungjinl@amazon.com Abstract There is an increasing trend in using neural methods for dialogue model evaluation. Lack of a framework to investigate these metrics can cause dialogue models to reflect their bi- ases and cause unforeseen problems during in- teractions. In this work, we propose an ad- versarial test-suite which generates problem- atic variations of various dialogue aspects, e.g. logical entailment, using automatic heuristics. We show that dialogue metrics for both open- domain and task-oriented settings are biased in their assessments of different conversation behaviors and fail to properly penalize prob- lematic conversations, by analyzing their as- sessments of these problematic examples. We conclude that variability in training method- ologies and data-induced biases are some of the main causes of these problems. We also conduct an investigation into the metric behav- iors using a black-box interpretability model which corroborates our findings and provides evidence that metrics pay attention to the prob- lematic conversational constructs signaling a misunderstanding of different conversation se- mantics. 1 Introduction Automatic evaluation of natural language models in general and dialogue models in specific has been a focus of ongoing research. The gold standard for evaluation of dialogues is human judgement (Meena et al., 2014; Ultes et al., 2013; Jang et al., 2020; Shim et al., 2021; Khalid et al., 2020b; Pan- fili et al., 2021) but human judgements are hard to obtain. Other than human judgements, dialogue simulations are used to judge different aspects of a model behavior (Jung et al., 2009; Eckert et al., 1997; Cuayáhuitl et al., 2010; Khalid et al., 2020a; Kreyssig et al., 2018; Sun et al., 2021a). Neural models of dialogue rely on text similarity metrics like BLEU, ROUGE or METEOR (Papineni et al., 2002; Lin, 2004; Banerjee and Lavie, 2005), how- ever these do not correlate well with the human A: Do you like singing? B: Yes, I do. A: Let’s go to a ETV tonight. B: But I can’t sing it well. A: It doesn’t matter. No one will laugh at you. B: OK, I’Il go with you. When shall we go? Actual Response: What about six? Adversarial Response: OK. Let’s make it. Human Score: 0.05; Adversarial Score: 0.43 Table 1: This example presents a problematic judge- ment by the DialogRPT metric. This metric scores an irrelevant response higher than a relevant and natural response. This example highlights the volatile nature of neural evaluation metrics. judgement (Lowe et al., 2017). Recent research focuses on the use of neural net- works to tackle this problem (Kreyssig et al., 2018; Sun et al., 2021a; Lowe et al., 2017; Jang et al., 2020; Shim et al., 2021; Mehri and Eskenazi, 2020a; Pang et …

Content truncated to 500 words