Logo Lanfrica

hangyav/anchor-embeddings

Domaine:

natural language processing

Type de record:

model
Créateur:
han
Hôte:
Anchor-based cross-lingual word embeddings for very low-resource languages. # Anchored bi- and multilingual word embeddings This project aims at building cross-lingual word embeddings for low-resource languages, which lack large amounts of monolingual data. Instead of building monolingual word embeddings for multiple languages and aligning them in two independent steps, it builds the target language embeddings in a single step by *anchoring* them to the embeddings space of a high resource language. Both bilingual and multilingual embeddings are supported. For further details see our published papers. ## Released Resources * Embeddings built in our latest paper) for languages listed below: **download** * English (eng) * Kazakh (kaz) * Tagalog (tgl) * Icelandic (ice) * Swahili (swa) * Chuvas (chv) * Yakut (sah) * Faroese (fao) * Hiligaynon (hil) ## Setup ```bash pip install -r requirements.txt # put MUSE under the ./MUSE directory ``` NOTE: Developed with python version 3.8.18. ## Running the Experiments To run the experiments, use the following command: ```bash python3 run_experiment.py ``` or to save the output log and results to a file ```bash python3 run_experiment.py 2>&1 | tee ``` Replace ` ` with the path to your JSON configuration file. For more information regarding the JSON configuration files see the documentation under the experiments directory. JSON configuraions can be built manually, or generated using `build_chain_setup.py`. For further details see ``` python3 build_chain_setup.py -h ``` ## Troubleshooting If you encounter any issues while running the experiments, here are a few things you can try: - Ensure that all the paths in the JSON configuration file are correct and that the files exist. - Make sure that you have the necessary permissions to read the files and write to the directories specified in the JSON configuration file. - If you're getting out-of-memory errors, try reducing the `vector_count` or using a machine with more memory. If you're still having issues, please open an issue on the pr …

Languages

Licenses