Judge the Judges: A Large-Scale Evaluation Study of Neural Language Models for Online Review Generation

By Cristina Garbacea, Samuel Carton, Shiyan Yan, Qiaozhu Mei. 2019

34 citations recorded by the source corpus.

We conduct a large-scale, systematic study to evaluate the existing evaluation methods for natural language generation in the context of generating online product reviews. We compare human-based evaluators with a variety of automated evaluation procedures, including discriminative evaluators that measure how well machine-generated text can be distinguished from human-written text, as well as word overlap metrics that assess how similar the generated text compares to human-written references. We determine to what extent these different evaluators agree on the ranking of a dozen of state-of-the-art generators for online product reviews. We find that human evaluators do not correlate well with discriminative evaluators, leaving a bigger question of whether adversarial accuracy is the correct objective for natural language generation. In general, distinguishing machine-generated text is challenging even for human evaluators, and human decisions correlate better with lexical overlaps. We find lexical diversity an intriguing metric that is indicative of the assessments of different evaluators. A post-experiment survey of participants provides insights into how to evaluate and improve the quality of natural language generation systems.

Subfields: Artificial Intelligence, General Social Sciences, Political Science and International Relations, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In General Social Sciences

In Political Science and International Relations

In Evaluation & Benchmarks

Citations in the corpus

All That’s ‘Human’ Is Not Gold: Evaluating Human Evaluation of Generated Text (2021, methodology): “This task format, inspired by the Turing (1950) Test, is used to compare the quality of machine-generated text to human-authored text and, as models’ fluency improves, to analyze NLG models’ ability to “fool” readers (Garbacea et al., 2019; Ippolito et al., 2020; Brown et al., 2020).”
Learning to Compare for Better Training and Evaluation of Open Domain Natural Language Generation Models (2020, background): “Recent study shows that the discriminator accuracy does not correlate well with human preference (Garbacea et al. 2019).”
Evaluation of Text Generation: A Survey (2020, background): “, 2017), machine translation (Goutte, 2006), online review generation (Garbacea et al., 2019), interactive systems (Hastie & Belz, 2014a), and conversational dialog systems (Deriu et al.”
With Little Power Comes Great Responsibility (2020, background): “We did find one single case of authors performing a power analysis to estimate sample size among the papers we surveyed (Garbacea et al., 2019).”
A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications (2020, methodology): “Furthermore, GANs have been widely utilized in other NLP applications such as question answer selection [393], [394], poetry generation [395], talent-job fit [396], and review detection and geneneration [397], [398].”
UNION: An Unreferenced Metric for Evaluating Open-ended Story Generation (2020, methodology): “However, training such a discriminator can be easily over-fitted to a specific dataset, thereby leading to poor generalization and low correlation with human judgments (Garbacea et al., 2019).”
CTRLEval: An Unsupervised Reference-Free Metric for Evaluating Controlled Text Generation (2022, unknown): “This may cause over-fitting to task-specific data and degrade the generalization ability of metrics (Garbacea et al., 2019), thereby giving unstable evaluation of generated texts from different models or with different qualities (Guan and Huang, 2020).”
The Perils of Using Mechanical Turk to Evaluate Open-Ended Text Generation (2021, background): “Previous studies have asked workers to choose from (Mori et al., 2019) or distinguish between human-written and machine-generated texts (Garbacea et al., 2019; Ippolito et al., 2020; Clark et al., 2021).”
A Survey on Generative Adversarial Networks: Variants, Applications, and Training (2020, background): “, speech, and voice applications, such as Utility-GAN [156] and Multi-Scale Matching GAN (MSM-GAN) [157] for question-answer selection, Resume-GAN [158] for talent-job fit, Natural Language GAN (NLG) [159] and Fake-GAN [160] for review detection and generation, Information Retrieval GAN (IR-GAN) [16”