Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On Bilingual Lexicon Induction with Large Language Models

Yaoyiran Li, Anna Korhonen, Ivan Vulić

TL;DR

This work investigates bilingual lexicon induction (BLI) using multilingual text-to-text large language models (mLLMs). It systematically evaluates zero-shot and few-shot prompting, augmented by nearest-neighbor in-context examples, across 18 open-source mLLMs from five families on XLING and PanLex-BLI. The key finding is that few-shot prompting with in-context neighbor examples often yields the best BLI performance, achieving new state-of-the-art results on XLING, particularly with LLaMA-13B, while lower-resource languages still lag behind CLWE-based methods. The study also explores template design, retrieval augmentation, and limited BLI-oriented fine-tuning, highlighting both the promise and the current limitations of prompting-based BLI for broad language coverage and practical deployment.

Abstract

Bilingual Lexicon Induction (BLI) is a core task in multilingual NLP that still, to a large extent, relies on calculating cross-lingual word representations. Inspired by the global paradigm shift in NLP towards Large Language Models (LLMs), we examine the potential of the latest generation of LLMs for the development of bilingual lexicons. We ask the following research question: Is it possible to prompt and fine-tune multilingual LLMs (mLLMs) for BLI, and how does this approach compare against and complement current BLI approaches? To this end, we systematically study 1) zero-shot prompting for unsupervised BLI and 2) few-shot in-context prompting with a set of seed translation pairs, both without any LLM fine-tuning, as well as 3) standard BLI-oriented fine-tuning of smaller LLMs. We experiment with 18 open-source text-to-text mLLMs of different sizes (from 0.3B to 13B parameters) on two standard BLI benchmarks covering a range of typologically diverse languages. Our work is the first to demonstrate strong BLI capabilities of text-to-text mLLMs. The results reveal that few-shot prompting with in-context examples from nearest neighbours achieves the best performance, establishing new state-of-the-art BLI scores for many language pairs. We also conduct a series of in-depth analyses and ablation studies, providing more insights on BLI with (m)LLMs, also along with their limitations.

On Bilingual Lexicon Induction with Large Language Models

TL;DR

This work investigates bilingual lexicon induction (BLI) using multilingual text-to-text large language models (mLLMs). It systematically evaluates zero-shot and few-shot prompting, augmented by nearest-neighbor in-context examples, across 18 open-source mLLMs from five families on XLING and PanLex-BLI. The key finding is that few-shot prompting with in-context neighbor examples often yields the best BLI performance, achieving new state-of-the-art results on XLING, particularly with LLaMA-13B, while lower-resource languages still lag behind CLWE-based methods. The study also explores template design, retrieval augmentation, and limited BLI-oriented fine-tuning, highlighting both the promise and the current limitations of prompting-based BLI for broad language coverage and practical deployment.

Abstract

Bilingual Lexicon Induction (BLI) is a core task in multilingual NLP that still, to a large extent, relies on calculating cross-lingual word representations. Inspired by the global paradigm shift in NLP towards Large Language Models (LLMs), we examine the potential of the latest generation of LLMs for the development of bilingual lexicons. We ask the following research question: Is it possible to prompt and fine-tune multilingual LLMs (mLLMs) for BLI, and how does this approach compare against and complement current BLI approaches? To this end, we systematically study 1) zero-shot prompting for unsupervised BLI and 2) few-shot in-context prompting with a set of seed translation pairs, both without any LLM fine-tuning, as well as 3) standard BLI-oriented fine-tuning of smaller LLMs. We experiment with 18 open-source text-to-text mLLMs of different sizes (from 0.3B to 13B parameters) on two standard BLI benchmarks covering a range of typologically diverse languages. Our work is the first to demonstrate strong BLI capabilities of text-to-text mLLMs. The results reveal that few-shot prompting with in-context examples from nearest neighbours achieves the best performance, establishing new state-of-the-art BLI scores for many language pairs. We also conduct a series of in-depth analyses and ablation studies, providing more insights on BLI with (m)LLMs, also along with their limitations.
Paper Structure (18 sections, 2 figures, 19 tables)

This paper contains 18 sections, 2 figures, 19 tables.

Table of Contents

  1. Introduction and Motivation
  2. Related Work
  3. Methodology
  4. Prompting Multilingual LLMs for BLI
  5. Retrieval-Augmented In-Context Learning
  6. Template Design and BLI Inference
  7. BLI-Oriented Fine-Tuning
  8. Experimental Setup
  9. Results and Discussion
  10. Main Results
  11. Further Analyses
  12. Further Discussion
  13. Conclusion
  14. Languages
  15. Reproducibility Checklist
  16. ...and 3 more sections

Figures (2)

  • Figure 1: Averaged BLI score versus model size ($0.3$B to $13$B): (left) $|\mathcal{D}_S|$=$5$K; (middle) $|\mathcal{D}_S|$=$1$K; (right) $|\mathcal{D}_S|$=0.
  • Figure 2: BLI scores averaged over 20 BLI directions from XLING with respect to the number of in-context examples $n$ ($0$ to $10$), with mT5$_{\text{large}}$ and LLaMA$_{13\text{B}}$ in both $5$K and $1$K BLI setups.