MGTV-AI/Mango-SubBench

Benchmark

Mango-SubBench

Audiovisual subtitle translation evaluation prioritizes practical delivery standards for global long-form videos rather than mere similarity to human references. Qualified subtitles require simultaneous compliance with semantic accuracy, natural expression, standardized segmentation, valid timestamps and traceable quality, which cannot be fully assessed by single automatic metrics. To mitigate this limitation, we collaborate with Beijing International Studies University to build a dedicated audiovisual translation benchmark equipped with high-quality datasets and an automated-centric evaluation framework.It supports scalable, reproducible comprehensive model evaluation, with supplementary random manual sampling to assess subtle subtitle characteristics including character tone, cultural adaptability and viewing experience. Meanwhile, in collaboration with Beijing International Studies University, we are establishing specialized evaluation criteria and a dataset for audiovisual subtitle translation. We will open-source this benchmark in due course and keep updating model performance metrics evaluated on it in the near future. See GitHub for more details.

dataset

The evaluation set is constructed based on real-world film and television subtitle scenarios, covering 11 languages with a total of 8019 sentence-level samples.

Language	Sample Count
Indonesian(印尼语)	1077
English(英语)	762
Vietnamese(越南语)	657
Malay(马来语)	489
Thai(泰语)	457
Korean(韩语)	699
Japanese(日语)	665
Arabic(阿语)	388
French(法语)	1026
Spanish(西语)	523
Russian(俄语)	1026
All	8019

Criteria

To enable horizontal comparison of different models and translation batches on the same evaluation set, we derive a weighted composite score with following metrics:

1. Semantic similarity： Semantic similarity quantifies the semantic proximity between model translations and reference translations based on text embeddings.

2. COMET： As a neural evaluation model, COMET integrates source text, human reference translation and machine translation output to predict the consistency between translations and human quality judgments. It is suitable for measuring whether the semantics of the source text are accurately preserved in translations.

3. BLEU-2： BLEU-2 provides supplementary information on local surface matching and phrase consistency.

Performance

We evaluate our Mango-MT-9B against three commercial large models (DeepSeek-v4-pro, Gemini-3-pro, GPT-5.4) across 11 languages with five core translation metrics: Semantic similarity, BLEU-2, COMET On translation benchmark, Mango-MT outperforms Gemini, DeepSeek and GPT across all 11 evaluated languages. This demonstrates that our model delivers powerful multilingual translation capabilities for film and drama content, and possesses significant commercial deployment value.

Evaluation Results on S(overall score) are:

Models	Malay	Thai	English	Indonesian	Vietnamese	Russian	French	Japanese	Korean	Spanish	Arabic
DeepSeek-V4-Pro	0.76	0.80	0.85	0.75	0.85	0.76	0.79	0.80	0.70	0.82	0.78
Gemini-3-Pro	0.78	0.88	0.86	0.79	0.87	0.79	0.82	0.82	0.73	0.83	0.78
GPT-5.4	0.75	0.83	0.83	0.76	0.84	0.77	0.79	0.81	0.71	0.81	0.73
Mango-MT	0.86	0.91	0.90	0.93	0.94	0.89	0.90	0.87	0.81	0.92	0.81