Near-Negative Distinction: Giving a Second Life to Human Evaluation Datasets

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2094 - 2108 December 7-11, 2022 ©2022 Association for Computational Linguistics Near-Negative Distinction: Giving a Second Life to Human Evaluation Datasets Philippe Laban Chien-Sheng Wu Wenhao Liu Caiming Xiong Salesforce AI Research {plaban, wu.jason, wenhao.liu, cxiong}@salesforce.com Abstract Precisely assessing the progress in natural lan- guage generation (NLG) tasks is challenging, and human evaluation to establish a preference in a model’s output over another is often nec- essary. However, human evaluation is usually costly, difficult to reproduce, and non-reusable. In this paper, we propose a new and simple automatic evaluation method for NLG called Near-Negative Distinction (NND) that repur- poses prior human annotations into NND tests. In an NND test, an NLG model must place a higher likelihood on a high-quality output can- didate than on a near-negative candidate with a known error. Model performance is established by the number of NND tests a model passes, as well as the distribution over task-specific er- rors the model fails on. Through experiments on three NLG tasks (question generation, ques- tion answering, and summarization), we show that NND achieves a higher correlation with hu- man judgments than standard NLG evaluation metrics. We then illustrate NND evaluation in four practical scenarios, for example per- forming fine-grain model analysis, or studying model training dynamics. Our findings suggest that NND can give a second life to human anno- tations and provide low-cost NLG evaluation. 1 Introduction Pre-training of large language models has fueled recent progress in many natural language genera- tion (NLG) tasks such as summarization (Zhang et al., 2020), question answering (Tafjord and Clark, 2021) (Ng et al., 2019), and question gener- ation (Murakhovs’ka et al., 2022). However, quan- tifying this progress remains a challenge due to the open-ended nature of NLG. The gold standard for NLG evaluation is manual expert annotation: it can be highly precise and fully customized to an NLG task, helping identify model limitations, and setting the direction of future work. The main limitation of manual expert annotation is the complexity and cost associated with running 1 Human Annotated Candidates No Error Sev: 0 2 Not Fluent Sev: 1 4 No Error Sev: 0 5 Not Fluent Sev: 1 3 Not Factual Sev: 1 P( )= 0.6 Administer NND Tests 1 2 0.6 > 0.2 1 0.6 < 0.7 1 0.6 > 0.4 3 5 4 2 0.3 > 0.2 4 0.3 < 0.7 4 0.3 < 0.4 3 5 NND Results Overall Pass Rate Disfluent Not Factual 50% (3/6) 75% (3/4) 0% (0/2) Evaluated Model Candidate Likelihood 1 P( )= 0.2 P( )= 0.7 2 3 P( )= 0.3 P( )= 0.4 4 5 Error Category Pass Rate Figure 1: Stages of the Near-Negative Distinction framework for NLG evaluation. A pre-existing human evaluation is repurposed into a series of NND tests. an evaluation. The cost often increases linearly or quadratically with the number of models compared, restricting evaluation to a small number of models. To circumvent the …

Content truncated to 500 words