We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.CL

Change to browse by:

cs

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

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

Abstract: Precisely assessing the progress in natural language generation (NLG) tasks is challenging, and human evaluation to establish preference in a model's output over another is often necessary. 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 repurposes prior human annotations into NND tests. In an NND test, an NLG model must place higher likelihood on a high-quality output candidate 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 errors the model fails on. Through experiments on three NLG tasks (question generation, question answering, and summarization), we show that NND achieves higher correlation with human judgments than standard NLG evaluation metrics. We then illustrate NND evaluation in four practical scenarios, for example performing fine-grain model analysis, or studying model training dynamics. Our findings suggest NND can give a second life to human annotations and provide low-cost NLG evaluation.
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:2205.06871 [cs.CL]
  (or arXiv:2205.06871v1 [cs.CL] for this version)

Submission history

From: Philippe Laban [view email]
[v1] Fri, 13 May 2022 20:02:53 GMT (519kb,D)

Link back to: arXiv, form interface, contact.