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

Yaoyiran Li, Fangyu Liu, Ivan Vulić, and Anna Korhonen. 2022. Improving Bilingual Lexicon Induction with Cross-Encoder Reranking. In Findings of the Association for Computational Linguistics: EMNLP 2022. [arXiv]

BLICEr is a post-hoc reranking method that works in the synergy with any given Cross-lingual Word Embedding (CLWE) space to improve Bilingual Lexicon Induction (BLI) / Word Translation. BLICEr is applicable to any existing CLWE induction method such as ContrastiveBLI, RCSLS, and VecMap. Our method first 1) creates a cross-lingual word similarity dataset, comprising positive word pairs (i.e., true translations) and hard negative pairs induced from the original CLWE space, and then 2) fine-tunes an mPLM (e.g., mBERT or XLM-R) in a Cross Encoder manner to predict the similarity scores. At inference, we 3) combine the similarity score from the original CLWE space with the score from the BLI-tuned cross-encoder.

As reported in our paper, BLICEr is tested in four different BLI setups:

• Supervised, 5k seed translation pairs

• Semi-supervised, 1k seed translation pairs

• Unsupervised, 0 seed translation pairs

• Zero-shot, 0 translation pairs directly between source and target languages but assume seed pairs between them and a third language respectively (no overlapping)

• PyTorch >= 1.10.1
• Transformers >= 4.15.0
• Python >= 3.9.7
• Sentence-Transformers >= 2.1.0

Following ContrastiveBLI, our data are obtained from the XLING (8 languages, 56 BLI directions in total) and PanLex-BLI (15 lower-resource languages, 210 BLI directions in total); please refer to ContrastiveBLI for data preprocessing details.

Our BLICEr is compatible with any CLWE backbones. For brevity, our demo here is based on the state-of-the-art ContrastiveBLI 300-dim C1 CLWEs, which is derived with purely static fastText embeddings (ContrastiveBLI also provides even stronger 768-dim C2 CLWEs which are trained with both fastText and mBERT). Please modify the input/output directories accordingly when using different CLWEs.

python run_all.py

Output: source->target and target->source P@1 scores for each of λ values in [0, 0.01, 0.02, ... , 0.99, 1.0].

Please cite our paper if you find BLICEr useful. If you like our work, please ⭐ this repo.

@inproceedings{li-etal-2022-improving-bilingual,
    title     = {Improving Bilingual Lexicon Induction with Cross-Encoder Reranking},
    author    = {Li, Yaoyiran and Liu, Fangyu and Vuli{\'c}, Ivan and Korhonen, Anna},
    booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2022},
    year      = {2022}
}