| Challenge: | Recent embedding-based neural text evaluation metrics are not reliable due to black-box nature of neural networks. |
| Approach: | They propose to find a single adversarial text in the discrete space that is consistently evaluated as high-quality regardless of the test cases. |
| Outcome: | The proposed method outperforms translations generated individually for each source sentence in English-to-Japanese and English- to-German translation tasks. |
Similar Papers
The Inside Story: Towards Better Understanding of Machine Translation Neural Evaluation Metrics (2023.acl-short)
Copied to clipboard
| Challenge: | Neural metrics for machine translation evaluation are considered "black boxes" lexical overlap-based metrics are popular for evaluation of translation systems and algorithms . |
| Approach: | They develop and compare several neural explainability methods to understand translation errors . they aim to better understand the correspondence between token-level explanations and human annotated error spans . |
| Outcome: | The proposed methods leverage token-level information that can be directly attributed to translation errors. |
Robustness Tests for Automatic Machine Translation Metrics with Adversarial Attacks (2023.findings-emnlp)
Copied to clipboard
| Challenge: | BERTScore, BLEURT, and COMET are automatic evaluation metrics that are often underperformed on adversarially-synthesized texts. |
| Approach: | They examine MT evaluation metric performance on adversarially-synthesized texts . they validate that automatic metrics tend to overpenalize adversarial-degraded translations . |
| Outcome: | The results show that automatic metrics tend to overpenalize adversarially-degraded translations. |
Identifying Weaknesses in Machine Translation Metrics Through Minimum Bayes Risk Decoding: A Case Study for COMET (2022.aacl-main)
Copied to clipboard
| Challenge: | Neural metrics have a high correlation with human judgements but they are hard to eliminate due to their "black box" nature. |
| Approach: | They propose to use minimum bayes risk decoding to explore and quantify weaknesses in COMET models. |
| Outcome: | The proposed model is not sensitive enough to discrepancies in numbers and named entities, and is hard to remove by training on additional synthetic data. |
COMET: A Neural Framework for MT Evaluation (2020.emnlp-main)
Copied to clipboard
| Challenge: | Historically, metrics for evaluating the quality of machine translation (MT) have relied on basic, lexical-level features such as counting the number of matching n-grams between the MT hypothesis and the reference translation. |
| Approach: | They propose a neural framework for training multilingual machine translation evaluation models which exploits human judgements to obtain new state-of-the-art levels of correlation with MT quality. |
| Outcome: | The proposed framework achieves state-of-the-art performance on the WMT 2019 Metrics shared task and demonstrate robustness to high-performing systems. |
One Single Hub Text Breaks CLIP: Identifying Vulnerabilities in Cross-Modal Encoders via Hubness (2026.acl-long)
Copied to clipboard
| Challenge: | et al., 2010) show that hub embeddings are close to many unrelated examples in high-dimensional embeddable spaces . cross-modal encoders that project different modalities into a shared space are useful for cross-module applications . |
| Approach: | They propose a method for identifying the hub embedding and its corresponding hub text . they use images to evaluate cross-modal encoders that project different modalities into a shared space . |
| Outcome: | The proposed method can identify a single hub embedding and its corresponding hub text . it achieves comparable or higher similarity scores than human-written reference captions in many images . |
Tangled up in BLEU: Reevaluating the Evaluation of Automatic Machine Translation Evaluation Metrics (2020.acl-main)
Copied to clipboard
| Challenge: | Existing methods for judging metrics are sensitive to the translations used for evaluation, leading to falsely confident conclusions about a metric’s efficacy. |
| Approach: | They propose a method for thresholding performance improvement under an automatic metric against human judgements by using a pairwise system ranking method. |
| Outcome: | The proposed method allows quantification of type I versus type II errors incurred, i.e., insignificant human differences in system quality that are accepted, and significant human differences that are rejected. |
MENLI: Robust Evaluation Metrics from Natural Language Inference (2023.tacl-1)
Copied to clipboard
| Challenge: | Recent proposed BERT-based evaluation metrics for text generation are vulnerable to adversarial attacks, e.g., relating to information correctness. |
| Approach: | They propose to use BERT-based evaluation metrics for text generation to evaluate text for semantic similarity but are vulnerable to adversarial attacks using Natural Language Inference. |
| Outcome: | The proposed metrics outperform existing summarization metrics but perform below SOTA MT metrics on standard benchmarks. |
BLEURT Has Universal Translations: An Analysis of Automatic Metrics by Minimum Risk Training (2023.acl-long)
Copied to clipboard
| Challenge: | Automated metrics are used for machine translation, but they are often considered to be black boxes with potential biases that are difficult to detect. |
| Approach: | They analyze automatic metrics from the perspective of their guidance for machine translation training. |
| Outcome: | The proposed measures improve the performance of machine translation models. |
Beyond N-Grams: Rethinking Evaluation Metrics and Strategies for Multilingual Abstractive Summarization (2025.acl-long)
Copied to clipboard
| Challenge: | n-gram-based metrics are considered indicative (even if imperfect) of human evaluation for English, but their suitability for other languages remains unclear. |
| Approach: | They systematically assess evaluation metrics for generation for languages and tasks using n-gram-based and neural-based metrics. |
| Outcome: | The proposed evaluation suite is based on eight languages from four typological families and shows that it is sensitivity to the language type at hand. |
DEMETR: Diagnosing Evaluation Metrics for Translation (2022.emnlp-main)
Copied to clipboard
| Challenge: | BLEU scores are based on string overlap, but they are opaque in comparison to newer learned metrics. |
| Approach: | They propose a dataset to evaluate MT evaluation metrics based on linguistic perturbations in English . they find learned metrics perform substantially better than string-based metrics . |
| Outcome: | The proposed dataset shows that learned metrics perform better than string-based metrics . the dataset contains 31K English examples that cover 35 different linguistic phenomena . |