SAIL-BLI

This repository is the official PyTorch implementation of the following paper:

Yaoyiran Li, Anna Korhonen, and Ivan Vulić. 2024. Self-Augmented In-Context Learning for Unsupervised Word Translation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024). [Paper]

SAIL aims to improve unsupervised Word Translation / Bilingual Lexicon Induction (BLI) by (1) inferring a high-confidence word translation dictionary with zero-shot prompting, (2) then optionally refining the high-confidence dictionary iteratively with few-shot prompting where the in-context examples are from the high-confidence dictionary in the previous iteration, and (3) finally conducting evaluation on the BLI test set with few-shot prompting also deriving in-context samples from the latest high-confidence dictionary. The whole process does not leverage any ground-truth word translation pairs for training/few-shot learning and improves the BLI scores by typically 10 ~ 15 Precision@1 points on our BLI benchmarks comparing to zero-shot prompting.

Dependencies

• PyTorch>=1.10.1
• Transformers>=4.28.1
• OpenAI>=1.28.1 (optional)

Data

Following our previous work ContrastiveBLI, BLICEr and Prompt4BLI, our data is obtained from the XLING (8 languages, 56 BLI directions in total) and PanLex-BLI (15 lower-resource languages, 210 BLI directions in total).

Get XLING data:

sh get_xling_data.sh

For PanLex-BLI, please see ./get_panlex_data, where we provide the code for deriving the monolingual word embeddings.

Run the Code

Prepare BLI Vocabulary:

python run_extract_vocabularies.py

Run BLI Evaluation with SAIL (define key hyper-parameters, directories and language pairs to evaluate manually in run_bli.py):

python run_bli.py

(Optional) Run Zero-Shot Prompting Baseline Introduced in Prompt4BLI with LLaMA models:

python run_zero_shot.py

(Optional) Run Zero-Shot Prompting Baseline Introduced in Prompt4BLI with ChatGPT models:

python run_zero_shot_chatgpt.py

LLMs Used in Our Work

Main Experiments (SAIL):

Our main experiments include four pretrained LLaMA models without instruction-tuning.

LLM	(Hugging Face) Model ID
LLaMA-7B	"huggyllama/llama-7b"
LLaMA-13B	"huggyllama/llama-13b"
LLaMA2-7B	"meta-llama/Llama-2-7b-hf"
LLaMA2-13B	"meta-llama/Llama-2-13b-hf"

Update: LLaMA3-8B meta-llama/Meta-Llama-3-8B is also supported in our code repo now.

Additional ChatGPT Experiments (only Zero-Shot Prompting, corresponding to our paper's Section 4.2):

Since the instruction-tuning of ChatGPT models is likely to cover large-scale parallel data for machine translation, they are not suitable for unsupervised BLI (in addition, the RLHF can also include supervision signals of word/sentence-level translation from annotators/users). We report ChatGPT results derived with zero-shot prompting only as a reference.

LLM	(OpenAI API) Model ID
GPT-3.5	"gpt-3.5-turbo-0125"
GPT-4	"gpt-4-turbo-2024-04-09"

Self-Augmented High-Confidence Dictionaries

We also release the self-augmented dictionaries derived with LLAMA2-13B as discussed in our paper's section 4.2 at ./AugmentedDicts-LLAMA2-13B. These high-confidence dictionaries are inferred with N_it = 1, N_f = 5000, and with word back-translation.

Citation

Please cite our paper if you find SAIL-BLI useful.

@inproceedings{li-etal-2024-self-augmented,
    title     = {Self-Augmented In-Context Learning for Unsupervised Word Translation},
    author    = {Li, Yaoyiran and Korhonen, Anna and Vuli{\'c}, Ivan},
    booktitle = {Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics},    
    year      = {2024}
}