Aligning More Words with High Precision for Small Bilingual Corpora

Sur-Jin Ker, Jason J. S. ChangCOLING 1996

TL;DR

This paper introduces "SenseAlign," a novel word alignment algorithm that leverages class-based rules derived from thesauri to achieve high precision and broad coverage, even with small bilingual corpora. Unlike traditional word-by-word statistical methods, SenseAlign resolves word sense ambiguity and offers a more efficient and robust approach to aligning words in bilingual texts.

What Problem Does This Paper Address?

Aligning words in bilingual corpora is a fundamental task for many Natural Language Processing (NLP) applications, including machine translation, bilingual lexicography, and word sense disambiguation. Traditional statistical machine translation (SMT) approaches often rely on word-by-word translation probabilities, which require massive amounts of training data (hundreds of thousands of sentences). This reliance on large datasets leads to a significant problem: less frequent words or words with multiple translations often lack sufficient statistical evidence for confident alignment, resulting in incomplete or incorrect alignments. Furthermore, these methods struggle with word sense ambiguity, where a single word can have different meanings and thus different translations depending on the context. The paper aims to address this gap by developing an algorithm that can achieve high-precision word alignment with significantly smaller training corpora and inherently handle word sense ambiguity.

Methods

The core of SenseAlign is a class-based word alignment system that utilizes both existing and automatically acquired lexical knowledge. Instead of relying on direct word-by-word translation probabilities, SenseAlign employs class-based rules. These rules are acquired from bilingual materials, such as example sentences and their translations or definition sentences from machine-readable dictionaries.

Key components and features of SenseAlign include:

  • Greedy Alignment Algorithm: This algorithm makes decisions by selecting preferred connections based on a composite score of several factors: applicability, specificity, fan-out, relative distortion probabilities, and evidence from bilingual dictionaries.
  • Lexical Preprocessing: This involves morphological analysis, part-of-speech tagging, and idiom identification for both languages. It also includes morpho-syntactic analyses to handle language-specific structures, aiming to bring sentences closer in word count.
  • Thesauri for Classification: The system uses two thesauri to classify words. This classification allows words to align with target words based on the collective translation tendency of words within the same class. Class-based rules are noted to have fewer parameters, be easier to acquire, and be more broadly applicable.
  • Class-Based Rule Acquisition: Rules can be learned either from bilingual corpora (example sentences and translations) or from definition sentences in a machine-readable dictionary. These definitions are disambiguated using sense divisions based on thesauri.
  • Similarity Factor: The algorithm incorporates similarity between connection targets and dictionary translations, exploiting the "thesauri effect" where words with common characters might be synonyms.
  • Relative Distortion: Instead of absolute position distortion, SenseAlign uses relative distortion, which accounts for the tendency of translation to preserve contiguous syntactical structures.

Datasets

The primary dataset used for training and evaluation is the Longman English-Chinese Dictionary of Contemporary English (LecDOCE).

  • Training Data: Approximately 25,000 bilingual example sentences from LecDOCE were used to acquire rules and empirically determine probability functions for various factors.
  • Testing Data:
    • Inside Test: Fifty sentence pairs from LecDOCE.
  • Outside Test: 416 sentence pairs from a book on English sentence patterns, containing 55 sets of typical sentence patterns. This set was noted to contain more common words, making them easier to align.

Key Insights and Ideas

  • Thesauri are Powerful for Alignment: Leveraging class-based rules derived from thesauri significantly broadens the coverage of word alignment while maintaining high precision, even with limited training data.
  • Class-Based Rules Reduce Complexity: By grouping words into classes, the algorithm reduces the number of candidates to consider, making the alignment process more efficient and increasing the likelihood of finding correct translations.
  • Word Sense Ambiguity Resolution: The class-based approach inherently helps in resolving word sense ambiguity by aligning words based on their contextual senses as defined by the thesauri.
  • Small Corpora are Viable: The algorithm demonstrates that effective word alignment can be achieved with corpora that are significantly smaller than those typically required by traditional word-by-word statistical methods.
  • Synergy of Lexical Resources: Combining machine-readable dictionaries and thesauri with corpora creates a more robust and generalizable word alignment system.

Key Results

The paper reports impressive results, particularly when compared to baseline methods:

  • High Coverage and Precision: SenseAlign aligns over 80% of word pairs with a comparable high precision rate (93%) even when trained on a corpus one-tenth the size of that used in Gale and Church (1991).
  • Comparison with DictAlign:
    • Inside Test: Full SenseAlign achieved 81.8% coverage and 93.3% precision, significantly outperforming DictAlign (which achieved up to 39.2% coverage and 82.1% precision with similarity filtering).
  • Outside Test: Full SenseAlign achieved 84.7% coverage and 93.4% precision, again significantly outperforming DictAlign (up to 41.1% coverage and 85.7% precision).
  • Effectiveness of Similarity: Exploiting the "thesauri effect" (similarity between words) in DictAlign could increase coverage by nearly three-fold but at the cost of a 10% decrease in precision. SenseAlign, by incorporating this and other factors, achieves both high coverage and precision.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that class-based models might be over-generalized and could miss word-specific rules.
  • While the approach is general, the current implementation is specific to English-Chinese corpora.
  • Future Work:
    • The authors suggest that the algorithm's performance can be further improved by enhancing its various components, such as morphological analyses, bilingual dictionaries, monolingual thesauri, and rule acquisition.
  • They believe the linguistic issues addressed are general enough to adapt the algorithm to other language pairs, with Japanese being a particularly promising candidate.
  • The proposed algorithm can produce effective word-alignment results with sense tagging, which can serve as a basis for other NLP tasks like word sense disambiguation and PP attachment.

Summary updated

BibTeX
@inproceedings{ker1996aligning,
  author = {Sur-Jin Ker and Jason J. S. Chang},
  title = {{Aligning More Words with High Precision for Small Bilingual Corpora}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1996},
}
Figure