CroCoAlign: A Cross-Lingual, Context-Aware and Fully-Neural Sentence Alignment System for Long Texts.
This is the official repository for CroCoAlign: A Cross-Lingual, Context-Aware and Fully-Neural Sentence Alignment System for Long Texts.
This work has been published at EACL 2024 (main conference). If you use any part, please consider citing our paper as follows:
@inproceedings{molfese-etal-2024-neuralign,
title = "CroCoAlign: A Cross-Lingual, Context-Aware and Fully-Neural Sentence Alignment System for Long Texts",
author = "Molfese, Francesco and
Bejgu, Andrei and
Tedeschi, Simone and
Conia, Simone and
Navigli, Roberto",
editor = "Graham, Yvette and
Purver, Matthew",
booktitle = "Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = mar,
year = "2024",
address = "St. Julian{'}s, Malta",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.eacl-long.135",
pages = "2209--2220",
abstract = "Sentence alignment {--} establishing links between corresponding sentences in two related documents {--} is an important NLP task with several downstream applications, such as machine translation (MT). Despite the fact that existing sentence alignment systems have achieved promising results, their effectiveness is based on auxiliary information such as document metadata or machine-generated translations, as well as hyperparameter-sensitive techniques. Moreover, these systems often overlook the crucial role that context plays in the alignment process. In this paper, we address the aforementioned issues and propose CroCoAlign: the first context-aware, end-to-end and fully neural architecture for sentence alignment. Our system maps source and target sentences in long documents by contextualizing their sentence embeddings with respect to the other sentences in the document. We extensively evaluate CroCoAlign on a multilingual dataset consisting of 20 language pairs derived from the Opus project, and demonstrate that our model achieves state-of-the-art performance. To ensure reproducibility, we release our code and model checkpoints at https://github.com/Babelscape/CroCoAlign.",
}- Sentence alignment using sentence embeddings and context encoder.
- Support for various languages.
- Customizable alignment strategies.
- Evaluation metrics for alignment quality.
To install CroCoAlign, follow these steps:
- Clone the repository:
git clone https://github.com/Babelscape/CroCoAlign.git - Create a new conda environment from the env.yml file:
conda env create -f env.yml - Activate the environment:
conda activate crocoalign
You can download the official checkpoint at the following link.
Under the src/sentence_aligner/preprocessing folder you can find two python scripts.
- dataset_generator.py is needed to convert the original xml Opus documents into the jsonl format required for training. Under the data folder you can already find the preprocessed data for convenience.
- precompute_embeddings.py (OPTIONAL) can be used to precompute the embeddings for the data.
-
Set the variable core.data_dir contained in the conf/default.yml file to the path containing the data for train, validation and test.
-
If you want the system to compute embeddings at runtime, set the variables conf.nn.data.precomputed_embeddings and conf.nn.module.precomputed_embeddings to False. If you have run the precompute_embeddings.py script to generate embeddings for the data, you can set both the variables to True in order to let the system skip the sentence embedding step during the forward pass.
-
Set the variables conf.transformer_name and conf.tokenizer_transformer_name to the desired sentence transformers using the HuggingFace name (e.g. "sentence-transformers/LaBSE") or a local path.
-
To train a new instance of CroCoAlign, run the command:
PYTHONPATH="src" python src/sentence_aligner/run.py param_1 ... param_n
Where param_1 ... param_n are the parameters of the network that can be modified at runtime.
You can consult which parameters can be changed by accessing the conf directory and its subdirectories.
Alternatively, you can also modify the parameter directly in the .yaml files instead of modifying them at runtime during training.
To evaluate CroCoAlign on the Opus book dataset, run the following command:
PYTHONPATH="src" python src/sentence_aligner/evaluate.py ckpt_path test_data_path
You can call the script with the -h command to get information about the available command options.
The data is already splitted into train, val and test and its available under the data folder.
To reproduce the paper results against Vecalign, you can run the following script:
python results/scripts/paper_results.py results/data/opus/books/crocoalign-{version}.tsv
Where {version} needs to be replaced with one of the available recovery strategies.
See the content of the results/data/opus/books/ folder to see the available options.
To align your own parallel documents using CroCoAlign, you can run the following command:
PYTHONPATH="src" python src/sentence_aligner/crocoalign.py source_document target_document
You can call the script with the -h command to get information about the available command options.
The default format of the source and target document is considered to be .txt.
In case you would like to try another sentence encoder, you can either provide a .jsonl files containing the source and target sentences with or without precomputed sentence embeddings (by using the -p command) or selecting another sentence encoder at runtime.
You can also select the desired output format of the final alignment (either .tsv or .jsonl).
CroCoAlign is licensed under the CC BY-SA-NC 4.0 license. The text of the license can be found here.
We underline that the dataset we used in our experiments has been extracted from the Opus website, which was introduced in the following work: J. Tiedemann, 2012, Parallel Data, Tools and Interfaces in OPUS. In Proceedings of the 8th International Conference on Language Resources and Evaluation (LREC 2012).
The code in this repository is built on top of .
The icon appearing in this README and in the official paper title was taken from Flaticon.com.