Research map / Large language models / Language, safety and interpretability
FiRE: Fine-grained Ranking Evaluation for Machine Translation
Wenyang Gao, Yinghao Yang, Xi Jin, Jing Li, Yue Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use criterion-conditioned FiRE aggregation instead of a direct overall ranking prompt (ablation) | DeepSeek-R1 · Overall pairwise ranking EN→ZH | agreement with human (%): 64.3 -> 65.3 | Improved |
| use criterion-conditioned FiRE aggregation instead of a direct overall ranking prompt (ablation) | DeepSeek-R1 · Overall pairwise ranking RU→ZH | agreement with human (%): 66.7 -> 70.1 | Improved |
| use LLM pairwise FiRE instead of error-based evaluators on fluency and style | MT evaluation · Consistency of style EN→ZH | agreement with human (%): 3.0 -> 61.4 (GEMBA-MQM vs FiRE) | Improved |
| use LLM-based FiRE instead of trained regression and ranking metrics | MT evaluation · Overall pairwise ranking EN→ZH | agreement with human (%): 61.6 -> 65.3 (MetricX-24-XXL vs FiRE) | Improved |
| evaluate harder examples (two of three annotators agree) rather than easy ones (ablation) | LLM evaluators · Faithfulness EN→ZH | agreement with human (%): 57.6 -> 43.4 | Hurt |
| swap candidate order for dedicated ranker (ablation) | MT-Ranker-XXL · Overall RU→ZH | agreement with human (%): 61.6 -> 41.9 | Hurt |
| swap candidate order with FiRE aggregation (ablation) | DeepSeek-R1-FiRE · Overall EN→ZH | agreement with human (%): 65.3 -> 65.9 | No clear effect |
| reconcile judgments at criterion level before aggregation versus collapsing each order to overall decision first (ablation) | FiRE · Position bias mitigation | agreement with human | Improved |
| add locale convention as a fourth criterion (FiREplus) | FiRE · 250 new EN→ZH pairs | overall agreement | Improved |
| add FiREplus criterion on full benchmark | FiRE · EN→ZH and RU→ZH full benchmark | overall agreement | No clear effect |
| evaluate with same-lineage LLM judge | GPT-4o and Qwen2.5-72B evaluators · EN→ZH and RU→ZH | self-preference | Hurt |
| use reasoning models (DeepSeek-R1, QwQ-32B) as evaluators | FiRE · Position consistency EN→ZH | position consistency | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MT-Ranker-XXL — MT-Ranker: Reference-free machine translation evaluation by inter-system ranking (ICLR 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MT-Ranker: Reference-free machine translation evaluation by inter-system ranking (ICLR 2024 · similarity 0.94)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.88)
- Language Ranker: A Metric for Quantifying LLM Performance Across High and Low-Resource Languages (AAAI 2025 · similarity 0.88)
- Are Large Reasoning Models Good Translation Evaluators? Analysis and Performance Boost (NeurIPS 2025 · similarity 0.88)
- Sentence-level Aggregation of Lexical Metrics Correlates Stronger with Human Judgements than Corpus-level Aggregation (AAAI 2025 · similarity 0.87)
- Learning from others' mistakes: Finetuning machine translation models with span-level error annotations (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
