How NOT To Evaluate Your Dialogue System: An Empirical Study of Unsupervised Evaluation Metrics for Dialogue Response Generation
Open accessThe DOI matches its Crossref record; 47 of the 47 references it cites were found in a registry or an index.
Abstract
We investigate evaluation metrics for dialogue response generation systems where supervised labels, such as task completion, are not available.Recent works in response generation have adopted metrics from machine translation to compare a model's generated response to a single target response.We show that these metrics correlate very weakly with human judgements in the non-technical Twitter domain, and not at all in the technical Ubuntu domain.We provide quantitative and qualitative results highlighting specific weaknesses in existing metrics, and provide recommendations for future development of better automatic evaluation metrics for dialogue systems.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
In the citing paper
“For high-variation domains like our generation tasks, on the other hand, wordoverlap-based machine translation metrics like BLEU (Post 2018) are expected to have low correlation with human quality judgments (Liu et al. 2016).”
From Can Large Language Models Transform Computational Social Science? (2024) · cites this paper as Liu et al. 2016
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Hochreiter, S., & Schmidhuber, J. (1997) · Neural Computation · doi:10.1162/neco.1997.9.8.1735 · Cited by 102,318
Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation
Cho, K., van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., & Bengio, Y. (2014) · Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.3115/v1/d14-1179 · Cited by 25,545 · cited 1×
“, 2015), machine translation (Cho et al., 2014), and dialogue response generation systems (Sordoni et al.”
BackgroundDistributed Representations of Words and Phrases and their Compositionality
Mikolov, T., Sutskever, I., Chen, K., Corrado, G., & Dean, J. (2013) · arXiv · doi:10.48550/arxiv.1310.4546 · Cited by 24,533
BLEU: a method for automatic evaluation of machine translation
Papineni, K., Roukos, S., Ward, T., & Zhu, W.-J. (2001) · Proceedings of the 40th Annual Meeting on Association for Computational Linguistics - ACL '02 · doi:10.3115/1073083.1073135 · Cited by 21,344
(2004) · Cited by 9,718
Weighted kappa: nominal scale agreement provision for scaled disagreement or partial credit
Cohen, J. (1968) · Psychological Bulletin · doi:10.1037/h0026256 · Cited by 8,736 · cited 2×
“2 w.r.t. the other respondents, which is a standard measure for inter-rater agreement (Cohen, 1968).”
BackgroundLandauer, T. K., & Dumais, S. T. (1997) · Psychological Review · doi:10.1037/0033-295x.104.2.211 · Cited by 6,282
(2005) · Cited by 4,152
Generating Sequences With Recurrent Neural Networks
Graves, A. (2013) · arXiv · doi:10.48550/arxiv.1308.0850 · Cited by 2,894
A Diversity-Promoting Objective Function for Neural Conversation Models
Li, J., Galley, M., Brockett, C., Gao, J., & Dolan, B. (2016) · Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies · doi:10.18653/v1/n16-1014 · Cited by 1,926
Cited by
Items 101–117 of 117
On the Generation of Medical Question-Answer Pairs
Shen, S., Li, Y., Du, N., Wu, X., Xie, Y., Ge, S., Yang, T., Wang, K., Liang, X., & Fan, W. (2020) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v34i05.6410 · Cited by 25
Sequence to Sequence Modeling for User Simulation in Dialog Systems
Crook, P., & Marin, A. (2017) · Interspeech 2017 · doi:10.21437/interspeech.2017-161 · Cited by 25
“[21] discuss the ineffectiveness of other unsupervised metrics, such as BLEU [22], METEOR [23], and ROUGE [24], in particular in the context of evaluating the dialog system responses which may or may not have strong surface form overlap with the ground truth.”
BackgroundD-Score: Holistic Dialogue Evaluation Without Reference
Zhang, C., Lee, G., D'Haro, L. F., & Li, H. (2021) · IEEE/ACM Transactions on Audio, Speech, and Language Processing · doi:10.1109/taslp.2021.3074012 · Cited by 21
Am I Me or You? State-of-the-Art Dialogue Models Cannot Maintain an Identity
Shuster, K., Urbanek, J., Szlam, A., & Weston, J. (2022) · Findings of the Association for Computational Linguistics: NAACL 2022 · doi:10.18653/v1/2022.findings-naacl.182 · Cited by 20
Abu-Jeyyab, M., Alrosan, S., & Alkhawaldeh, I. (2023) · High Yield Medical Reviews · doi:10.59707/hymrfbya5348 · Cited by 17
Spurious Correlations in Reference-Free Evaluation of Text Generation
Durmus, E., Ladhak, F., & Hashimoto, T. (2022) · Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2022.acl-long.102 · Cited by 16
CHBias: Bias Evaluation and Mitigation of Chinese Conversational Language Models
Zhao, J., Fang, M., Shi, Z., Li, Y., Chen, L., & Pechenizkiy, M. (2023) · Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2023.acl-long.757 · Cited by 14
Grounding as a Collaborative Process
Benotti, L., & Blackburn, P. (2021) · Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume · doi:10.18653/v1/2021.eacl-main.41 · Cited by 14
“For human evaluation, it is not enough to show a few turns to the annotators (Liu et al., 2016).”
BackgroundJang, J., Boo, M., & Kim, H. (2023) · Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/2023.emnlp-main.838 · Cited by 13
(2021) · arXiv (Cornell University) · Cited by 13
Towards User-Centric Text-to-Text Generation: A Survey
Yang, D., & Flek, L. (2021) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-83527-9_1 · Cited by 10
Chen, D., Chen, H., Yang, Y., Lin, A., & Yu, Z. (2021) · arXiv · doi:10.48550/arxiv.2104.00783 · Cited by 6
“BLEU score) tend to become unreliable in these situations (Liu et al., 2016).”
BackgroundEnhancing Code Generation via Bidirectional Comment-Level Mutual Grounding
Di, Y., & Zhang, T. (2025) · arXiv · doi:10.48550/arxiv.2505.07768 · Cited by 3
Chen, N., Li, H., Huang, J., Wang, B., & Li, J. (2024) · arXiv · doi:10.48550/arxiv.2402.11975 · Cited by 3
“Maintaining long-term conversations has always been a long-standing pursuit in current open-domain dialogue systems (Liu et al., 2016; Zhang et al., 2018; Kann et al., 2022), commonly known as chatbots or conversational agents.”
Preference Leakage: A Contamination Problem in LLM-as-a-judge
Li, D., Sun, R., Huang, Y., Zhong, M., Jiang, B., Han, J., Zhang, X., Wang, W., & Liu, H. (2025) · arXiv · doi:10.48550/arxiv.2502.01534 · Cited by 134
“To address these challenges, the paradigm of LLM-as-a-judge [79, 34, 24, 82, 37] has been proposed, designed to leverage LLM as evaluators to assess response quality.”
Chen, Y., Zhang, C., Luo, D., D'Haro, L. F., Tan, R. T., & Li, H. (2024) · arXiv · doi:10.48550/arxiv.2405.14646
Kim, J., Chay, W., Hwang, H., Kyung, D., Chung, H., Cho, E., Kwon, Y., Jo, Y., & Choi, E. (2024) · arXiv · doi:10.48550/arxiv.2406.13144
Metadata
- Authors
- Liu, C.-W., Lowe, R., Serban, I., Noseworthy, M., Charlin, L., & Pineau, J.
- Venue
- Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing
- Published
- 2016
- Pages
- 2122-2132
- OpenAlex
- W2963903950
- Cited by
- 1,228
- References
- 47
- FWCI
- 80.59
- Type
- conference paper
- Language
- en
- Open access
- hybrid
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.