PR + RQ ≈ PQ: Transliteration Mining Using Bridge Language

AAAI 2010

TL;DR

This paper introduces BRIDGE-CCA, a novel method for "Transliteration Mining" that learns to identify equivalent names across languages (like English and Tamil) even when direct parallel name data is unavailable. It cleverly uses a "bridge language" (like Kannada) and Canonical Correlation Analysis (CCA) to create a shared feature space, achieving performance comparable to methods that do require direct parallel data.

What Problem Does This Paper Address?

In many cross-language applications like Machine Translation and Cross-Language Information Retrieval (CLIR), correctly handling proper names is crucial. Names often appear as "out-of-vocabulary" words, and mis-translating them degrades performance. "Transliteration Mining" aims to automatically discover pairs of names that are transliterations of each other across languages.

The core challenge this paper tackles is the "resource-poor" scenario for transliteration mining. Traditional methods rely on training data consisting of parallel name pairs in the source and target languages (e.g., English-Tamil pairs). However, for many language pairs, such direct parallel data is scarce or non-existent. This paper addresses the gap by proposing a method that doesn't require direct parallel names in the target language pair (PQ), but instead leverages parallel data with a "bridge language" (R), where data like PR and RQ is more readily available.

Methods

The paper's core innovation lies in using Canonical Correlation Analysis (CCA) to learn a common geometric feature space for names across different languages.

  1. Canonical Correlation Analysis (CCA): CCA is a statistical technique used to find linear relationships between two sets of variables. In this context, it's used to project feature vectors of names from different languages into a shared, lower-dimensional space where similar names are mapped to nearby points.
  2. Bridge Language Approach:
    • Direct PQ Training (for context): If parallel names in PQ are available, CCA can be applied directly to feature vectors of names from P and Q to find projection matrices (A, B) that maximize the correlation between their projected representations in a common space.
  3. Bridge Language Training (Novelty): When PQ parallel names are not available, the authors propose using parallel names in PR and RQ. They construct composite feature vectors. For a PR pair (p, r) and an RQ pair (r, q), they create views: one based on the bridge language name (r) and another composite view that combines information from the source name (p) and the target name (q). CCA is then applied to these composite views to learn projection matrices (A, B1, B2) that map names from P, R, and Q into a common space.
  4. Similarity Measurement: Once names are projected into this common feature space, their similarity is computed as the Euclidean distance between their corresponding vectors. A threshold is then applied to determine if they are transliterations.

Datasets

The paper uses comparable corpora for training and evaluation. The specific datasets are not new but are the same ones used by prior work (Udupa et al., 2008, 2009b) to allow for direct comparison.

  • Training Data: Consists of parallel name pairs between English and Kannada (Eng-Kan), and Kannada and Tamil (Kan-Tam), Kannada and Hindi (Kan-Hin), and Kannada and Russian (Kan-Rus). Crucially, no direct English-Tamil, English-Hindi, or English-Russian parallel name pairs are used for training the BRIDGE-CCA model.
  • Test Data: Comparable corpora for English-Tamil (ET), English-Hindi (EH), and English-Russian (ER) across different environments: IDEAL, NEAR-IDEAL, and REAL.

Key Insights and Ideas

  • The Power of Bridge Languages: The most significant insight is that a readily available "bridge language" (like English, which is often resource-rich) can be leveraged to overcome the lack of direct parallel data for less resourced language pairs. This dramatically expands the applicability of transliteration mining.
  • CCA for Cross-Lingual Representation: Using CCA to learn a shared geometric feature space is a powerful technique for aligning representations of names across languages, even when their scripts are entirely different.
  • Composite Views for Bridge Learning: The clever construction of composite feature vectors that combine information from the source, bridge, and target languages is key to enabling CCA to learn the desired mappings without direct PQ data.
  • Language Agnosticism: The proposed method doesn't rely on language-specific rules or phonetic mappings, making it broadly applicable.

Key Results

  • Competitive Performance: BRIDGE-CCA achieved performance comparable to the state-of-the-art MINT system, which does use direct parallel PQ training data. In the IDEAL environment, BRIDGE-CCA achieved at least 88% of MINT's performance. In NEAR-IDEAL, it achieved 80-90%.
  • Surprising Superiority: In the REAL environment for English-Tamil, BRIDGE-CCA surprisingly outperformed MINT.
  • Computational Efficiency: The method is computationally efficient, taking less than two minutes to mine transliterations across all testbeds and data environments, compared to much longer times for some previous methods.
  • Robustness: When tested with direct PQ training data (though not the focus), BRIDGE-CCA's performance was within ±2% of a direct CCA classifier, highlighting the strength of CCA itself.

Limitations and Future Work

  • Small Amount of Direct Data: The authors suggest exploring how to incorporate a small amount of direct PQ parallel data into the model to potentially further improve performance.
  • Transliteration Generation: They also express interest in exploring Transliteration Generation (as opposed to mining) in this resource-poor scenario.
  • Temporal Alignment: While the method doesn't require temporal alignment of corpora, it's worth noting that some prior methods did, and this is a difference that might influence performance in certain real-world scenarios.

Summary updated

BibTeX
@inproceedings{2010pr,
  title = {{PR + RQ ≈ PQ: Transliteration Mining Using Bridge Language}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2010},
}
Figure