Automatic Bilingual Phrase Extraction from Comparable Corpora

Ahmet Aker, Yang Feng, Robert J. GaizauskasCOLING 2012 · OpenReview

Introduction

Statistical Machine Translation (SMT) systems are traditionally reliant on massive parallel corpora—identical texts translated across languages. However, for many language pairs and specialized domains, such resources are scarce. Researchers from the University of Sheffield have addressed this "data hunger" by developing a fully automated framework to extract parallel phrase pairs from "comparable corpora." Unlike parallel texts, comparable corpora consist of articles covering similar topics (e.g., international news) but written independently in different languages.

The "Generate and Test" Framework

The proposed system utilizes a two-stage approach to identify translation "needles" within the "haystack" of comparable data:

  1. Phrase Pair Generation: The system iterates through source and target sentences from document-aligned corpora, generating all possible phrase pairings within a length of 2 to 7 words.
  2. SVM Classification: A binary Support Vector Machine (SVM) classifier evaluates these pairs. It uses 21 distinct features to distinguish between "parallel" and "non-parallel" phrases.

Key features for the classifier include length differences, translation probabilities derived from GIZA++ dictionaries, and cognate-based similarity. For languages with different scripts, such as Greek, the researchers implemented a transliteration step to enable string similarity comparisons (e.g., Levenshtein distance) for named entities.

Experimental Setup

The framework was tested on three language pairs: English-German (EN-DE), English-Greek (EN-EL), and English-Latvian (EN-LV). Training data for the classifier was generated automatically by sampling "true" pairs from the JRC-Acquis parallel corpus and creating "false" pairs from unrelated phrases. The comparable data consisted of news articles aligned at the document level based on metadata like titles and dates.

Results and Impact

The classifier demonstrated high efficacy, achieving over 97% accuracy across all tested languages. The researchers specifically optimized for the $F_{0.5}$ metric, prioritizing precision over recall. This ensures that the extracted phrases are highly accurate, preventing "noise" from degrading the SMT model.

Table 3: Performance of the SVM classifier on test data, showing high precision and accuracy across language pairs.

The ultimate goal was to improve translation quality. By injecting the extracted phrase pairs into a baseline SMT system, the researchers observed significant improvements in BLEU scores. The most notable relative gains were seen in English-Latvian, highlighting the method's value for under-resourced languages where traditional parallel data is scarce.

Table 4: BLEU score improvements demonstrating the impact of injecting extracted phrases into the SMT system.

Discussion and Future Work

While successful, the system faced challenges with morphological complexity, particularly German compound words, which the current feature set struggled to capture. Additionally, the classifier occasionally misidentified phrases with identical word sets but different semantic orders as parallel.

Future work aims to incorporate sub-word morphology and explore "bootstrapping" techniques—using extracted phrases to build better dictionaries, which in turn can be used to extract even more phrases in an iterative cycle. This fully automated approach represents a significant step toward making SMT viable for a wider range of global languages.

Summary updated

BibTeX
@inproceedings{aker2012automatic,
  author = {Ahmet Aker and Yang Feng and Robert J. Gaizauskas},
  title = {{Automatic Bilingual Phrase Extraction from Comparable Corpora}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2012},
  url = {https://openreview.net/forum?id=3AGw8zkTjd},
}
Figure