A Corpus-Based Approach to Automatic Compound Extraction

Keh-Yih Su, Ming-Wen Wu, Jing-Shin ChangACL 1994 · DOI

Published in the mid-1990s, this research represents a foundational step in applying statistical methods to a classic challenge in Natural Language Processing (NLP): the automatic identification of multi-word "compounds" that function as single semantic units.

The Challenge of Compound Extraction

In Machine Translation (MT), compounds—such as "disk drive" or "operating system"—are notoriously difficult to handle. Many are non-compositional, meaning their meaning cannot be derived simply by translating their individual components. For instance, a "green house" (a house painted green) differs fundamentally from a "greenhouse" (a structure for plants).

Failure to recognize these compounds leads to translation errors and parsing ambiguity. While human lexicographers can manually update dictionaries, the rapid evolution of technical language makes manual maintenance unsustainable. This paper proposes an automated framework to "mine" these terms from large text corpora, significantly reducing manual labor.

Methodology: A Statistical and Syntactic Hybrid

The authors frame compound extraction as a two-class classification problem, determining whether an n-gram (a sequence of $n$ words) belongs to a "Compound" or "Non-Compound" cluster. Their approach utilizes a Likelihood Ratio Test based on three primary features:

  1. Mutual Information (MI): Measures the statistical association between words. High MI indicates words appear together far more often than chance would suggest.
  2. Relative Frequency Count (RFC): Measures the n-gram's prevalence in the text. High frequency suggests a standard domain term.
  3. Part-of-Speech (POS) Filtering: Recognizing that most compounds follow specific patterns (e.g., [Noun, Noun]), the authors used a probabilistic tagger to filter out syntactically improbable candidates.

A key insight of the study was that MI and RFC are largely uncorrelated. By combining them, the system effectively filters out "false alarms"—pairs that are frequent but coincidental (like "of the") or rare but highly associated (statistical flukes).

Experimental Setup

The study utilized technical documentation from computer manuals, a domain where compounds are highly prevalent. The training set consisted of approximately 192,440 words, while the testing set contained roughly 49,314 words. The research focused specifically on bigrams and trigrams, which constitute the vast majority of technical compounds.

Key Results and Discussion

The system proved exceptionally effective at capturing compounds, prioritizing recall to ensure lexicographers missed as few terms as possible.

Performance metrics for bigram extraction showing high recall across training and testing sets.

As shown in the performance data, the system achieved a recall rate of over 96% for both bigrams and trigrams. While precision hovered between 40% and 48%, this trade-off is advantageous for dictionary maintenance; it is far more efficient for a human to delete a few incorrect suggestions than to manually search a 50,000-word corpus for missing terms.

Qualitatively, the system's top extractions were highly accurate. In the top five extracted trigrams, four were valid technical compounds essential for a domain-specific dictionary.

Qualitative examples of successfully extracted bigrams and trigrams from technical manuals.

Conclusion

By layering syntactic constraints over statistical measures, the authors demonstrated that automated systems could effectively assist in the labor-intensive task of dictionary expansion. While limited by a "precision gap" and the exclusion of very low-frequency terms, the framework provided a significant reduction in the human effort required to keep machine translation systems current with evolving technical terminology.

Summary updated

BibTeX
@inproceedings{su1994corpus,
  author = {Keh-Yih Su and Ming-Wen Wu and Jing-Shin Chang},
  title = {{A Corpus-Based Approach to Automatic Compound Extraction}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1994},
  doi = {10.3115/981732.981765},
}
Figure