An Alignment Method for Noisy Parallel Corpora based on Image Processing Techniques

Jason S. Chang, Mathis H. M. ChenACL 1997 · DOI

Introduction

Bilingual alignment—the task of matching sentences and words between two languages—is a cornerstone of modern natural language processing. However, traditional methods often struggle with "noisy" corpora, where OCR errors, free translations, or significant deletions disrupt the linear progression of text. This paper introduces a novel approach that reimagines bitext alignment as a computer vision problem. By treating a bilingual corpus as a gray-level image, or "dotplot," the authors leverage image processing techniques to achieve robust alignment in challenging conditions.

The Challenge of Noisy Bitext

Standard alignment algorithms, typically based on Dynamic Programming (DP), rely on the assumption that the text follows a relatively consistent order. In real-world scenarios, such as scanned documents or non-literal translations, this assumption fails. A single large gap or a reordered paragraph can cause DP-based systems to lose track entirely. The goal of this research was to develop a method that remains resilient even when the data is sparse and filled with statistical noise.

Methodology: Text as an Image

The core of the proposed system, PlotAlign, is the transformation of linguistic data into a discrete gray-level image.

1. Lexical Translation Probability (LTP) Estimation

The process begins by calculating the likelihood that any given English word $x$ corresponds to a Chinese word $y$. This creates a "dotplot" where pixel intensity represents translation probability. As shown in the initial estimation, this raw data is extremely noisy, making it difficult to discern a clear path manually.

Initial LTP estimation showing high levels of noise in the bilingual mapping.

2. Convolution and Smoothing

To address local gaps and "missing pixels," the system applies a 5x5 convolution filter. This smoothing process uses local context to reinforce probable connections and fill in small voids where translation evidence might be weak, effectively "guessing" connections based on neighboring high-probability dots.

3. Texture Analysis

Valid alignments typically form dense regions within the image. By using projectional transformations, the system reduces the 2D image into 1D density features. This allows it to distinguish between the "signal" of a real translation and the "noise" of random word coincidences.

4. Global Path Detection via Hough Transform

To identify the overall trajectory of the translation, the authors employ the Hough Transform, a technique commonly used in computer vision to detect lines in noisy environments. This allows the system to find the "spine" of the translation across the entire corpus, even in the presence of massive deletions or inversions.

The final alignment path identified through image processing techniques, showing a clear diagonal trajectory despite initial noise.

Key Insights

The strength of this approach lies in mapping linguistic constraints to visual patterns:

  • Structure Preservation maps to Edges: Words appearing near each other in the source text should appear near each other in the target.
  • One-to-One Mapping maps to Texture: Most words have a primary counterpart, creating specific density patterns.
  • Monotonicity maps to Lines: Translations generally follow a linear path from the beginning to the end of the text.

Results and Robustness

The system was evaluated on a difficult "free" translation where only a fraction of the text had direct counterparts (40% of English and 70% of Chinese). Despite this, the combination of local convolution and global Hough Transform successfully suppressed noise and recovered the primary alignment path. While the Hough Transform excels at global consistency, the authors found that texture analysis was crucial for recovering shorter, isolated segments that might otherwise be ignored by the global line detector.

Conclusion

By shifting the paradigm from string matching to image processing, the authors demonstrate a highly robust method for aligning noisy parallel corpora. This vision-based approach provides a multi-resolution framework that gathers evidence at both local and global scales, making it significantly more resilient to the imperfections of real-world bilingual data.

Summary updated

BibTeX
@inproceedings{chang1997alignment,
  author = {Jason S. Chang and Mathis H. M. Chen},
  title = {{An Alignment Method for Noisy Parallel Corpora based on Image Processing Techniques}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1997},
  doi = {10.3115/976909.979655},
}
Figure