Automatic Noun Classification by Using Japanese-English Word Pairs

Naomi InoueACL 1991 · DOI

Introduction

In the field of Natural Language Processing (NLP), the automatic classification of nouns into semantic categories is a fundamental task. Traditionally, this process relied on manual labor to construct thesauri, a method that is both time-consuming and difficult to scale. Modern approaches often leverage the distributional hypothesis, which posits that words appearing in similar linguistic contexts—specifically, nouns that share the same governing verbs—are semantically related. However, this logic is frequently undermined by polysemy, where a single verb possesses multiple distinct meanings, leading to incorrect semantic groupings.

This paper presents a novel methodology to mitigate the polysemy problem by utilizing Japanese-English word pairs as a semantic filter. By incorporating translation data, the system can distinguish between different senses of a verb, thereby refining the accuracy of noun clustering.

The Polysemy Challenge

The primary obstacle in distribution-based classification is the ambiguity of verbs. For instance, the Japanese verb dasu can mean "to send" or "to publish." Without disambiguation, a system might treat these as the same context, incorrectly linking nouns associated with mailing to those associated with printing. This research argues that while a word may be ambiguous in its source language, its translation into a target language (English) often serves as a natural disambiguator.

Methodology: Translation as a Semantic Filter

The proposed method treats a Japanese-English verb pair (e.g., dasu-send vs. dasu-publish) as a unique "sense" of the verb. The classification process involves three main steps:

  1. Sense Disambiguation: Using bilingual word pairs to define distinct semantic contexts.
  2. Semantic Distance Calculation: A mathematical metric is applied to determine the distance between two nouns based on the ratio of shared verb senses to the total number of verb senses associated with them.
  3. Centroid Clustering: Hierarchical cluster analysis is performed using the centroid method. This specific clustering technique was chosen to prevent the "chain effect," where unrelated items are erroneously merged into a single large cluster.

Experimental Setup and Dataset

The study utilized the ATR Dialogue Database (ADD), which consists of approximately 150,000 words from simulated telephone and keyboard conversations regarding conference registrations and travel bookings. The experiment focused on 138 common nouns extracted from the most frequent words in the telephone dialogue task. These sentences were previously subjected to morphological analysis and assigned 63 deep semantic cases (e.g., agent, object, locative).

Results and Analysis

The performance of the bilingual word pair method was compared against a baseline method that used only raw Japanese verb forms.

Performance comparison (P) between the proposed word pair method and the standard verb form method as the number of nouns increases.

As shown in the performance graph, the bilingual method significantly outperformed the standard method when classifying smaller sets of nouns (fewer than 30). In these instances, the precision gained from distinguishing verb senses outweighed the noise inherent in the raw data. However, as the number of nouns increased, the performance gap narrowed, suggesting that high-frequency statistical patterns eventually begin to compensate for polysemous noise.

Qualitatively, the system produced highly coherent semantic clusters. For example, the system successfully grouped nouns related to documentation and physical materials.

Hierarchical clustering of nouns related to documents and materials, demonstrating semantic coherence.

In Figure 3, the dendrogram illustrates the successful grouping of terms such as "list" (リスト), "form" (用紙), "material" (資料), and "document" (書類). Similar success was found in clustering abstract concepts like "decision," "presentation," and "speech."

Conclusion and Future Work

The use of translation as a "free" semantic tagger provides an elegant solution to the polysemy problem in noun classification. While effective, the current model faces limitations regarding raw frequency weighting, where common but non-descriptive verbs (like "to be") can skew semantic distances. Future research will focus on automatically acquiring word pairs from raw bilingual corpora and developing normalization techniques to prioritize rare, highly descriptive verbs over generic ones.

Summary updated

BibTeX
@inproceedings{inoue1991automatic,
  author = {Naomi Inoue},
  title = {{Automatic Noun Classification by Using Japanese-English Word Pairs}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1991},
  doi = {10.3115/981344.981370},
}
Figure