Spanish-English Translation Model for the Scientific Domain

Description

This is a CTranslate2 Spanish-English translation model for the scientific domain, which uses the CA+OC+ES-EN OPUS-MT Transformer-Big (link) as its base model. It has been fine-tuned on a large parallel corpus with scientific texts, with special focus to the four pilot domains of the SciLake project:

Neuroscience
Cancer
Transportation
Energy

Dataset

The fine-tuning dataset consists of 4,145,412 EN-ES parallel sentences extracted from parallel theses and abstracts which have been acquired from multiple academic repositories.

Evaluation

We have evaluated the base and the fine-tuned models on 5 test sets:

Four which correspond to the pilot domains (Neuroscience, Cancer, Transportation, Energy) with each one containing 1,000 parallel sentences.
A general scientific which contains 3,000 parallel sentences from a wide range of scientific texts in other domains.

Model	Average of 4 domains			General Scientific
	SacreBLEU	chrF2++	COMET	SacreBLEU	chrF2++	COMET
Base	49.7	70.5	69.5	51	71.7	68.9
Fine-Tuned	51.9	71.7	70.9	54	73.1	71
Improvement	+2.2	+1.2	+1.4	+3	+1.4	+2.1

Usage

pip install ctranslate2 sentencepiece huggingface_hub

import ctranslate2
import sentencepiece as spm
from huggingface_hub import snapshot_download

repo_id = "ilsp/opus-mt-big-es-en_ct2_ft-SciLake"

# REPLACE WITH ACTUAL LOCAL DIRECTORY WHERE THE MODEL WILL BE DOWNLOADED
local_dir = ""

model_path = snapshot_download(repo_id=repo_id, local_dir=local_dir)

translator = ctranslate2.Translator(model_path, compute_type="auto")

sp_enc = spm.SentencePieceProcessor()
sp_enc.load(f"{model_path}/source.spm")

sp_dec = spm.SentencePieceProcessor()
sp_dec.load(f"{model_path}/target.spm")

def translate_text(input_text, sp_enc=sp_enc, sp_dec=sp_dec, translator=translator, beam_size=6):
    input_tokens = sp_enc.encode(input_text, out_type=str)
    results = translator.translate_batch([input_tokens],
                                         beam_size=beam_size,
                                         length_penalty=0,
                                         max_decoding_length=512,
                                         replace_unknowns=True)
    output_tokens = results[0].hypotheses[0]
    output_text = sp_dec.decode(output_tokens)
    return output_text
    
input_text = "La energía eléctrica es un insumo base de alta difusión, derivado de su capacidad para satisfacer todo tipo de necesidades."
translate_text(input_text)

# OUTPUT
# Electric power is a base input of high diffusion, derived from its ability to satisfy all types of needs.

Acknowledgements

This work was created within the SciLake project. We are grateful to the SciLake project for providing the resources and support that made this work possible. This project has received funding from the European Union’s Horizon Europe framework programme under grant agreement No. 101058573.