BLEU (bilingual evaluation understudy) is an automatic algorithm for evaluating the quality of text which has been machine-translated from one natural language to another.
COMET is also an automatic algorithm for evaluating the quality of translated text, based on neuronets.
From author: BLEU was introduced in 2002, and COMET in 2020. As for now, I recommend COMET scores more than BLEU.
Use this results just for reference.
Higher is better, no_translate can be used as baseline. Average on 100 examples from FLORES, offset = 150:
| fra->eng | eng->fra | rus->eng | eng->rus | |
|---|---|---|---|---|
| no_translate | 3.98 | 3.9 | 0.57 | 0.56 |
| libre_translate | 47.66 | 49.62 | 32.43 | 30.99 |
| fb_nllb_translate nllb-200-distilled-600M | 51.92 | 52.73 | 41.38 | 31.41 |
| fb_nllb_translate nllb-200-distilled-1.3B | 56.81 | 55 | 46.03 | 33.98 |
| fb_nllb_ctranslate2 JustFrederik/nllb-200-3.3B-ct2-float16 | 54.87 | 56.73 | 48.45 | 36.85 |
| fb_nllb_ctranslate2 JustFr-ik/nllb-200-distilled-1.3B-ct2-int8 | 56.12 | 56.45 | 46.07 | 34.56 |
| google_translate | 58.08 | 59.99 | 47.7 | 37.98 |
| deepl_translate | 57.67 | 59.93 | 50.09 | 38.91 |
| yandex_dev | ----- | ----- | 46.09 | 40.23 |
| openai_chat gpt-3.5-turbo (aka ChatGPT) | ----- | ----- | 41.49 | 30.9 |
| koboldapi_translate (alpaca7B-4bit) | 43.51 | 30.54 | 32 | 14.19 |
| koboldapi_translate (alpaca30B-4bit) | ----- | ----- | ----- | 24.0 |
| fb_mbart50 facebook/mbart-large-50-one-to-many-mmt | ----- | 48.79 | ----- | 28.55 |
| fb_mbart50 facebook/mbart-large-50-many-to-many-mmt | 50.26 | 48.93 | 42.47 | 28.56 |
| vsegpt_chat openai/gpt-3.5-turbo | ----- | ----- | 41.93 | 31.12 |
| vsegpt_chat openai/gpt-4 | ----- | ----- | 44.16 | 34.88 |
| vsegpt_chat anthropic/claude-instant-v1 | ----- | ----- | 41.88 | 29.67 |
| vsegpt_chat anthropic/claude-2 | 54.91 | 56.09 | 46.38 | 34.13 |
| opus_mt Helsinki-NLP/opus-mt-en-ru | ----- | ----- | ----- | 30.41 |
LLMs with errors:
- vsegpt_chat tiiuae/falcon-40b-instruct - a lot of fails
- vsegpt_chat google/palm-2-chat-bison - "I'm not able to help with that, as I'm only a language model."
- 'koboldapi_translate' on 'eng->rus' pair average BLEU score: 7.00: 80/100 on IlyaGusev-saiga_7b_lora_llamacpp-ggml-model-q4_1.bin, may be adjusting for input prompt needed
SOTA for opensource realization: multi_sources vsegpt_chat:lizpreciatior/lzlv-70b-fp16-hf,fb_nllb_ctranslate2 - this comparable to DeepL and Google Translate
Higher is better, no_translate2 can be used as baseline. Average on 100 examples from FLORES, offset = 150:
| fra->eng | eng->fra | rus->eng | eng->rus | |
|---|---|---|---|---|
| no_translate2 | 31.66 | 32.06 | 33.03 | 25.58 |
| no_translate | 79.2 | 70.19 | 69.3 | 44.82 |
| opus_mt Helsinki-NLP/opus-mt-en-ru | ----- | ----- | ----- | 82.22 |
| libre_translate | 86.66 | 82.36 | 80.36 | 83.34 |
| lingvanex | 87.92 | 86.99 | 84.75 | 86.3 |
| bloomz bigscience/bloomz-1b7 | 87.86 | 84.1 | ----- | ----- |
| vsegpt_chat recursal/eagle-7b | 87.18 | 83.67 | 84.56 | 75.94 |
| koboldapi_translate NikolayKozloff/ALMA-13B-GGUF | ----- | ----- | 84.64 | 87.92 |
| t5_mt utrobinmv/t5_translate_en_ru_zh_large_1024 | ----- | ----- | 86.05 | 86.53 |
| fb_nllb_translate nllb-200-distilled-1.3B | 89.01 | 87.95 | 86.91 | 88.57 |
| fb_nllb_ctranslate2 JustFrederik/nllb-200-3.3B-ct2-float16 | 88.74 | 88.32 | 87.25 | 88.83 |
| vsegpt_chat mistralai/mixtral-8x7b-instruct | 88.45 | 87.2 | 86.94 | 87.85 |
| vsegpt_chat lizpreciatior/lzlv-70b-fp16-hf | 88.69 | 87.17 | 86.91 | 88.15 |
| multi_sources vsegpt_chat:lizpreciatior/lzlv-70b-fp16-hf,fb_nllb_ctranslate2 | 89.14 | 88.22 | 87.22 | 89.87 |
| google_translate | 89.69 | 88.9 | 87.53 | 89.63 |
| deepl | 89.39 | 89.27 | 87.93 | 89.82 |
| vsegpt_chat anthropic/claude-instant-v1 | ----- | ----- | 85.73 | 88.13 |
| vsegpt_chat openai/gpt-3.5-turbo | ----- | ----- | 86.87 | 88.76 |
| vsegpt_chat openai/gpt-3.5-turbo-instruct | ----- | ----- | 85.23 | 87.46 |
| vsegpt_chat openai/gpt-4 | ----- | ----- | 87.02 | 89.54 |
| vsegpt_chat openai/gpt-4-1106-preview | ----- | ----- | ----- | 89.85 |
| vsegpt_chat openai/gpt-4-turbo | ----- | ----- | ----- | 89.76 |
| vsegpt_chat cohere/command-r-plus | ----- | ----- | ----- | 89.45 |
| vsegpt_chat anthropic/claude-2 | 89.27 | 89.17 | 87.47 | 89.85 |
| vsegpt_chat anthropic/claude-3-haiku | ----- | ----- | ----- | 89.5 |
| vsegpt_chat anthropic/claude-3-sonnet | ----- | ----- | ----- | 89.49 |
| vsegpt_chat anthropic/claude-3-opus | ----- | ----- | ----- | 90.75 |
| vsegpt_chat google/gemini-pro | ----- | ----- | ----- | 89.69 |
| multi_sources google_translate,deepl | 89.66 | 89.85 | 87.8 | 90.42 |
| multi_sources google_translate,deepl,vsegpt_chat* | 89.66 | 89.85 | 87.76 | 90.67 |
| yandex_dev | ----- | ----- | 87.34 | 90.27 |
| multi_sources google_translate,yandex_dev | ----- | ----- | 87.64 | 90.39 |
| multi_sources deepl,yandex_dev | ----- | ----- | 87.64 | 90.62 |
| multi_sources google_translate,deepl,yandex_dev | ----- | ----- | 87.74 | 90.63 |
| multi_sources google_translate,deepl,yandex_dev,vsegpt_chat* | ----- | ----- | 87.71 | 90.66 |
| multi_sources deepl,yandex_dev,vsegpt_chat* | ----- | ----- | 87.67 | 90.77 |
| multi_sources deepl,yandex_dev,vsegpt_chat** | ----- | ----- | ----- | 91.02 |
* vsegpt_chat with anthropic/claude-2 ** vsegpt_chat with anthropic/claude-3-opus
IMPORTANT: You interested how it will work on YOUR language pairs? It's easy, script already included, see "Automatic BLEU measurement" chapter.
Chain translation allow to translate phrases with mid-language (usually English)
BLEU scores
| jpn->rus | |
|---|---|
| no_translate | 0 |
| google_translate | 27.63 |
| deepl | 27.48 |
| use_mid_lang google_translate,deepl | 28.34 |
| use_mid_lang google_translate,google_translate | 27.62 |
| use_mid_lang deepl,deepl | 27.62 |
COMET scores
| jpn->rus | |
|---|---|
| no_translate | 56.85 |
| fb_nllb_ctranslate2 JustFrederik/nllb-200-3.3B-ct2-float16 | 86.46 |
| google_translate | 87.93 |
| deepl | 88.11 |
| use_mid_lang deepl->yandex_dev | 87.56 |
| use_mid_lang google_translate->deepl | 88.37 |
| use_mid_lang google_translate->google_translate | 87.67 |
| use_mid_lang deepl->deepl | 88.43 |
| multi_sources google_translate,deepl | 88.88 |
| use_mid_lang multi_sources->multi_sources* | 88.9 |
| multi_sources google_translate,deepl,use_mid_lang** | 89.05 |
- * multi_sources with "google_translate,deepl"
- ** use_mid_lang with "google_translate,deepl"
vsegpt_chat with anthropic/claude-2 get a lot of fails ("Can't understand", "Can't translate")
More results on different multi_sources settings:
| jpn->rus | |
|---|---|
| multi_sources google_translate,deepl,use_mid_lang:google_translate->deepl,use_mid_lang:google_translate->google_translate | 89.05 |
| multi_sources google_translate,deepl,use_mid_lang:google_translate->deepl,use_mid_lang:deepl->deepl | 89.15 |
| multi_sources google_translate,deepl,use_mid_lang:deepl->deepl | 89.04 |
| multi_sources**** | 89.25 |
**** model: google_translate,deepl,use_mid_lang:deepl->deepl,use_mid_lang:google_translate->deepl,use_mid_lang:google_translate->google_translate,use_mid_lang:deepl->google_translate
There are builded package to run BLEU and COMET estimation of plugin translation on different languages - so, you usually can reproduce our results.
There are pretty simple estimation based on FLORES dataset: https://huggingface.co/datasets/gsarti/flores_101/viewer
To estimate:
- install requirements-bleu.txt
- setup params in run_estimate_bleu.py (at beginning of file)
- run run_estimate_bleu.py
RECOMMENDATIONS:
- debug separate plugins first!
- To debug, use less BLEU_NUM_PHRASES.
Settings params:
# ----------------- key settings params ----------------
BLEU_PAIRS = "fra->eng,eng->fra,rus->eng,eng->rus" # pairs of language in terms of FLORES dataset https://huggingface.co/datasets/gsarti/flores_101/viewer
BLEU_PAIRS_2LETTERS = "fr->en,en->fr,ru->en,en->ru" # pairs of language codes that will be passed to plugin (from_lang, to_lang params)
BLEU_PLUGINS_AR = ["google_translate", "deepl", "multi_sources:google_translate,deepl"]
# plugins to estimate, array
# now you can run them in format "plugin:model", that works only if plugin support "on-the-fly" model change (usually YES for synthetic and online plugins, and NO for offline)
BLEU_NUM_PHRASES = 100 # num of phrases to estimate. Between 1 and 100 for now.
BLEU_START_PHRASE = 150 # offset from FLORES dataset to get NUM phrases
BLEU_METRIC = "bleu" # bleu | comet