A Discriminative Approach to Japanese Abbreviation Extraction

Naoaki Okazaki, Mitsuru Ishizuka, Jun'ichi TsujiiIJCNLP 2008 · OpenReview

Abstract

Naoaki Okazaki, Mitsuru Ishizuka, Jun’ichi Tsujii. Proceedings of the Third International Joint Conference on Natural Language Processing: Volume-II.

2008.

Introduction

Extracting abbreviations from text is a fundamental task in natural language processing (NLP). While English abbreviations often follow predictable patterns—such as acronyms where the short form's letters appear in the long form (e.g., "NASA" for "National Aeronautics and Space Administration")—Japanese abbreviations present a unique set of challenges. They frequently involve complex kanji combinations, phonetic shortenings, or translated aliases that do not adhere to simple character-matching rules.

Furthermore, parenthetical expressions in the form $X (Y)$ are ubiquitous but ambiguous. While some represent abbreviation definitions, many others denote attributes such as location or affiliation (e.g., "Beckham (England)"). This paper proposes a discriminative machine learning approach to distinguish true abbreviations from other parenthetical relations.

The Challenge of Japanese Abbreviations

The primary difficulty lies in the diversity of Japanese parenthetical expressions. As shown in Figure 1, "authentic" acronyms—where the abbreviation is formed directly from characters in the full name—account for only a tiny fraction (1.2%) of instances in the researchers' corpus. The majority of cases involve aliases, translated acronyms, or attributes that are not abbreviations at all. Relying on simple letter-matching rules would fail to capture the vast majority of Japanese shorthand.

Classification and examples of parenthetical expressions in Japanese text.

Methodology

The authors frame abbreviation extraction as a binary classification problem. Given a candidate pair $X (Y)$ extracted from text, a Support Vector Machine (SVM) classifier determines whether $Y$ is a valid abbreviation of $X$.

Key Features

  1. Paraphrase Ratio: The core innovation of this work. It relies on the behavioral intuition that an author introduces an abbreviation $Y$ to use it as a shorthand for $X$ later in the document. The feature measures the frequency of $Y$ relative to $X$ following the parenthetical definition, while ensuring $Y$ was not used prior to its introduction.
  2. Skew Divergence: A statistical measure of the similarity between the local linguistic contexts of $X$ and $Y$. If $Y$ is a true abbreviation, it should be interchangeable with $X$ in similar sentences.
  3. Co-occurrence Statistics: Standard metrics such as $\chi^2$ (Chi-square) and Log-likelihood ratio are used to assess the strength of the association between the two terms.
  4. Linguistic and Structural Features: These include part-of-speech (POS) tags, named entity (NE) tags, and character types (Kanji, Katakana, etc.).

Experimental Setup

The system was evaluated using a corpus of 7,887 parenthetical expressions extracted from the Mainichi Newspapers and Yomiuri Shimbun (1998–1999). The dataset contained 1,430 true abbreviations and 6,457 non-abbreviations, covering a wide range of types including aliases and translated acronyms.

Performance Evaluation

The SVM-based approach achieved an impressive accuracy of 95.7%, with 90.0% precision and 87.6% recall. A critical finding was the effectiveness of the "Paraphrase Ratio" feature. As illustrated in the Precision-Recall curves in Figure 6, the paraphrase rate significantly outperformed traditional statistical measures like co-occurrence frequency or skew divergence.

Precision-Recall curves for various features, highlighting the superiority of the paraphrase rate.

While linguistic features like character types were important for overall accuracy, the behavioral signal captured by the paraphrase ratio allowed the system to identify complex abbreviations that lack direct character overlap, such as "EU" for "欧州連合" (European Union).

Conclusion

This research demonstrates that a discriminative approach combining linguistic features with authorial behavior (paraphrase ratio) can highly effectively extract Japanese abbreviations. By moving beyond simple character-matching rules, the system handles the inherent complexity of non-alphabetic languages. Future work includes applying this methodology to other languages like Chinese and Korean, and adapting the system for the less structured environment of web documents.

Summary updated

BibTeX
@inproceedings{okazaki2008discriminative,
  author = {Naoaki Okazaki and Mitsuru Ishizuka and Jun'ichi Tsujii},
  title = {{A Discriminative Approach to Japanese Abbreviation Extraction}},
  booktitle = {IJCNLP},
  year = {2008},
  url = {https://openreview.net/forum?id=kfwHjdH74A9},
}

Loading related content...

Figure