Abstract
Named Entity Recognition (NER) is a vital component of information extraction, yet traditional systems often struggle with term mismatches when encountering previously unseen entities. This paper presents a novel NER scheme specifically designed for identifying restaurant names on the Korean web. By decomposing compound nouns and mapping them to multiple resolutions of a semantic tree (based on KorLex 1.5), the system leverages stable semantic categories rather than relying solely on surface-level text strings. Experimental results on a dataset of 49,153 nouns demonstrate that this multi-resolution approach achieves an average accuracy of 87.89%, significantly outperforming the 78.70% accuracy of baseline lexical systems.
Introduction
Identifying restaurant names in Korean text presents unique challenges compared to English. Korean lacks orthographic clues like capitalization, and restaurant names are frequently written as long compound nouns without spaces (e.g., "Leesangmandishhouse"). Traditional NER systems that rely on lexical matching—looking for specific strings seen during training—suffer from the "Out-of-Vocabulary" (OOV) problem, failing to recognize new businesses. To address this, we propose a method that focuses on the meaning of the components that constitute a restaurant name.
Methodology
The proposed system operates through a multi-stage pipeline designed to generalize the components of restaurant names:
- Multi-Word Decomposition: Using a morphological analyzer (KLT), compound nouns are broken into constituent units. For instance, a name meaning "Grandmother-Thick Broth-House" is split into three parts.
- Semantic Mapping: Each unit is mapped to a semantic hierarchy derived from Korean WordNet. This "multi-resolution" approach allows the system to view a word at various levels of abstraction, from specific leaf nodes (e.g., "Pizza") to broader categories (e.g., "Food").
- Transformation Functions: Features are extracted using top-down or bottom-up traversals of the tree, utilizing either fixed or variable windows to determine the depth of hierarchical information included in the feature vector.
- Classification: These semantic features are processed by a Support Vector Machine (SVM) to classify candidate nouns as restaurant names or non-restaurant entities.
Experimental Results
The system was tested on 49,153 nouns extracted from 467 web pages on Joyfood.com. The dataset was highly imbalanced, with restaurant names making up only 2.5% of the total nouns. Despite this, the multi-resolution approach showed robust performance.
As shown in the performance analysis, accuracy improves as the resolution increases. The system achieved its peak average accuracy of 87.89% using a top-down fixed window at the highest resolution (leaf level). This suggests that while the hierarchy provides a powerful framework for generalization during training, the specific semantic identity of the word remains the strongest signal for final classification.
Discussion and Conclusion
The core insight of this research is that while specific restaurant names are highly variable, the categories of words used to build them—such as Person, Location, Food, or Building types—are remarkably stable. By "blurring" definitions through a semantic hierarchy, the model can infer that if "Steak" is a clue for a restaurant, "Pizza" likely is as well, because both belong to the "Food" concept.
Future work will focus on addressing current limitations, such as WordNet coverage gaps (responsible for 9.12% of errors) and word sense ambiguity (7.10% of errors). Integrating Word Sense Disambiguation (WSD) could further refine the semantic mapping and improve performance in linguistically complex contexts.
Summary updated