Supervised Bilingual Lexicon Induction with Multiple Monolingual Signals

Ann Irvine, Chris Callison-BurchNAACL 2013

TL;DR

This paper introduces a supervised approach to bilingual lexicon induction, a task that has traditionally relied on unsupervised methods. By leveraging a small seed bilingual dictionary, the authors train a discriminative model that combines various monolingual signals, achieving a significant ~50% accuracy improvement over unsupervised baselines across 22 languages.

What Problem Does This Paper Address?

Bilingual lexicon induction is the process of finding word translation pairs between two languages using only their respective monolingual texts. While many existing methods rely on unsupervised techniques to combine signals like contextual similarity, temporal similarity, and topical information, this paper argues that supervised methods, which have proven superior in other machine learning tasks, have been surprisingly underutilized for this specific problem. The authors aim to bridge this gap by demonstrating that even a small amount of supervised data can dramatically improve lexicon induction performance, especially in low-resource settings where parallel data is scarce.

Methods

The core innovation of this paper is framing bilingual lexicon induction as a binary classification problem. For a given source language word, the model predicts whether a candidate target language word is its translation. This is achieved by:

  • Feature Engineering: Extracting various monolingual signals, including:
  • Contextual Similarity: Based on word co-occurrence in a window.
  • Temporal Similarity: Analyzing word appearance patterns over time in news articles.
  • Topical Similarity: Using inter-lingually linked Wikipedia pages to capture topic alignment.
  • Orthographic Similarity: Measuring edit distance between words (with transliteration for non-Roman scripts).
  • Frequency Similarity: Comparing the log frequencies of words in their respective monolingual corpora.
  • Supervised Learning: Training a linear classifier using stochastic gradient descent. The model is trained on a small seed bilingual dictionary, where known translation pairs serve as positive examples and random source-target word pairs are used as negative examples.
  • Feature Combination: The supervised model learns to optimally combine the diverse monolingual signals, assigning weights to each feature to make the final translation prediction.

Datasets

The authors utilize monolingual web crawl and Wikipedia data for 22 languages. The specific datasets include:

  • Web Crawled News Articles: Used for temporal and contextual similarity signals.
  • Wikipedia Pages: Used for topical similarity and contextual similarity. Inter-lingually linked pages are crucial for cross-lingual topic alignment.
  • Seed Bilingual Dictionary: A small set of known translation pairs is used for supervision (training, development, and testing). The paper does not introduce a new benchmark dataset but rather uses existing monolingual corpora and a small seed dictionary for evaluation.

Key Insights and Ideas

  • Supervision is Key: The paper strongly advocates for supervised learning in bilingual lexicon induction, showing it significantly outperforms unsupervised methods. This aligns with general trends in machine learning where supervised models excel.
  • Small Data, Big Gains: Surprisingly, only a few hundred translation pairs are sufficient to achieve strong performance, making this approach highly valuable for low-resource languages.
  • Monolingual Signals are Powerful: Even individually, signals derived from monolingual data (especially Wikipedia topic and context) are informative. However, their true power is unleashed when combined discriminatively.
  • Feature Importance Varies: The study highlights that Wikipedia topic and context signals are particularly strong predictors, while others like edit distance are more informative for certain language pairs (e.g., those with cognates).

Key Results

  • The supervised approach achieves an average relative gain in accuracy of nearly 50% over an unsupervised baseline.
  • The average top-10 accuracy across 22 languages for the unsupervised baseline is 30.4%, while the supervised method reaches 43.9%.
  • Large gains in accuracy are observed for all 22 languages investigated, regardless of their resource status.
  • The model's performance stabilizes after training with approximately 300 positive training instances, demonstrating the efficiency of the supervised approach.

Limitations and Future Work

  • Correlation vs. Causation: While an increase in monolingual data generally improves accuracy, the correlation isn't perfect. For instance, Urdu and Farsi show poorer performance despite large datasets, potentially due to smaller Wikipedia corpora which provide a strong topic signal.
  • Future Work: The authors suggest that their framework is flexible and can easily incorporate new monolingually-derived similarity metrics as additional features, paving the way for continuous improvement. The ability to add new signals without retraining the entire system is a significant advantage.

Summary updated

BibTeX
@inproceedings{irvine2013supervised,
  author = {Ann Irvine and Chris Callison-Burch},
  title = {{Supervised Bilingual Lexicon Induction with Multiple Monolingual Signals}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2013},
}
Figure