PR + RQ ALMOST EQUAL TO PQ: Transliteration Mining Using Bridge Language

Mitesh M. Khapra, Raghavendra Udupa, A. Kumaran, Pushpak BhattacharyyaAAAI 2010 · DOI

Introduction

Transliteration mining—the process of identifying name pairs across different scripts (e.g., "Mary" in English and "मारी" in Hindi)—is a critical component for modern Machine Translation (MT) and Cross-Language Information Retrieval (CLIR). These systems often fail when encountering "out-of-vocabulary" names. While supervised algorithms achieve high performance, they require thousands of parallel name pairs for every language combination. This paper introduces BRIDGE-CCA, a method that leverages a third "bridge" language to mine transliterations between two languages without any direct parallel training data.

The Resource-Poor Challenge

The primary bottleneck in transliteration mining is the dependency on supervised learning. For many language pairs, such as Greek and Tamil, the required parallel data simply does not exist. The authors propose a "bridge" strategy: if parallel data exists between language $P$ and bridge $R$, and between $R$ and language $Q$, can we mine pairs between $P$ and $Q$?

Methodology: BRIDGE-CCA

The core of the proposed approach is Canonical Correlation Analysis (CCA), a language-agnostic statistical method.

  1. Feature Representation: Names are decomposed into character bigram feature vectors (e.g., "Marie" $\rightarrow$ {ma, ar, ri, ie}).
  2. The Bridge Trick: To handle the absence of direct $P-Q$ data, the authors use "composite views." During training, they treat the bridge language ($R$) as one view and a combined vector of the source ($P$) and target ($Q$) as the second. When training on $P-R$ pairs, the $Q$ component is padded with zeros, and vice versa. This forces the bridge language to act as the mathematical "glue."
  3. Shared Geometric Space: The algorithm learns linear transformations that project all three languages into a single $k$-dimensional feature space.
Canonical Correlation Analysis (CCA) maps disparate feature spaces into a common k-dimensional space.
  1. Similarity Mining: In this shared space, the similarity between names is calculated as the Euclidean distance between their projected vectors. This method requires zero linguistic knowledge or manual "romanization" rules.

Experimental Setup

The researchers evaluated BRIDGE-CCA using English-Tamil, English-Hindi, and English-Russian pairs, primarily using Kannada as the bridge language. They tested the system across three data environments:

  • IDEAL: Perfectly aligned news articles.
  • NEAR-IDEAL: One-to-one article alignment exists but is hidden.
  • REAL: Noisy data from comparable corpora where direct article equivalents may not exist.

Notably, BRIDGE-CCA used only ~5,000 parallel pairs per link, compared to the ~16,000 pairs used by the supervised baseline.

Key Results

The performance was measured against MINT, a state-of-the-art supervised system with access to direct $P-Q$ training data.

Comparative performance (MRR) of BRIDGE-CCA versus the supervised MINT system.

As shown in the results table, BRIDGE-CCA is highly competitive. In "IDEAL" settings, it achieved 88% to 97.5% of the Mean Reciprocal Rank (MRR) of the fully supervised system. Most impressively, in the "REAL" noisy environment for English-Tamil, BRIDGE-CCA actually outperformed MINT (0.62 vs 0.58 MRR), demonstrating superior robustness despite the lack of direct training data.

Conclusion and Future Work

BRIDGE-CCA offers a powerful, computationally efficient (matching 1,000 words against 50,000 in under two minutes) solution for low-resource transliteration mining. While it depends on the availability of a suitable bridge language, its language-agnostic nature makes it widely applicable. Future research may explore extending this from mining existing lists to generating new transliterations, or incorporating small amounts of direct data to further boost performance.

Summary updated

BibTeX
@inproceedings{khapra2010pr,
  author = {Mitesh M. Khapra and Raghavendra Udupa and A. Kumaran and Pushpak Bhattacharyya},
  title = {{PR + RQ ALMOST EQUAL TO PQ: Transliteration Mining Using Bridge Language}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2010},
  doi = {10.1609/aaai.v24i1.7526},
}
Figure