Research map / Label-efficient and robust learning / Recognition and long-tail

Enhancing Bilingual Lexicon Induction via Bi-directional Translation Pair Retrieving

AAAI 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Multilingual language representation pretraining · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes bi-directional translation pair retrieval (BRTR/BRSR) for bilingual lexicon induction that re-scores source-to-target candidates using target-to-source retrieval to counter "source-side bias", improving P@1 across six language pairs in supervised and unsupervised settings.

Problem

Bilingual lexicon induction retrieves translations only source-to-target in cross-lingual word embeddings, producing asymmetric false-positive pairs that hurt precision.

Approach

Combine source-side nearest-neighbor scores with a target-side reverse retrieval score (rank-based BRTR or similarity-based BRSR) via a weighted sum with a tunable λ, so a candidate must also retrieve the source word back.

Key results

• BRSR improves average P@1 over baselines by 1.04 points (BRTR 0.62) in supervised BLI across six language pairs.
• With CLBLI, BRSR raises EN-FR (→) P@1 from 82.54 to 83.31 and EN-ZH average gains are notable; EN-ZH with MUSE rises from 25.00 to 33.60 P@1.
• In unsupervised BLI, VECMAP-BRSR improves average accuracy from 42.77 to 43.73 over the six pairs.

Measured effects reported in the paper

Surprising findings

• Restricting the k value of CSLS to match the proposed method's hyperparameters changes results only weakly, suggesting gains are not from neighborhood-size tuning.
• Unsupervised VECMAP achieves near-zero accuracy on EN-JA and EN-TH, and the gains on those pairs remain small.

Limitations

• Gains are modest on similar language pairs where baselines are already strong.
• Performance depends on λ, which must be tuned and varies with embedding quality and language similarity.
• Evaluated only with static fastText embeddings on MUSE; no contextual representations.

Experimental setup

Datasets: MUSE dataset (FastText 300-d embeddings, 200k vocab) MUSE test sets (1500 ground-truth pairs per language pair)

Benchmarks: MUSE BLI test sets (EN-FR, EN-ES, EN-IT, EN-ZH, EN-JA, EN-TH)

Metrics: P@1 P@5 P@10 accuracy

Baselines: MUSE VECMAP RCSLS LNMAP L1-refinement CLBLI

Builds on: CSLS VECMAP MUSE RCSLS CLBLI

Closest papers by meaning

• Bilingual Lexicon Induction from Non-Parallel Data with Minimal Supervision (AAAI 2017 · similarity 0.89)
• Filtered Inner Product Projection for Crosslingual Embedding Alignment (ICLR 2021 · similarity 0.87)
• Generalizing and Improving Bilingual Word Embedding Mappings with a Multi-Step Framework of Linear Transformations (AAAI 2018 · similarity 0.87)
• Cross-lingual Alignment vs Joint Training: A Comparative Study and A Simple Unified Framework (ICLR 2020 · similarity 0.87)
• CL2CM: Improving Cross-Lingual Cross-Modal Retrieval via Cross-Lingual Knowledge Transfer (AAAI 2024 · similarity 0.83)
• Constrained Decoding for Cross-lingual Label Projection (ICLR 2024 · similarity 0.83)