TL;DR
This paper presents a system for automatically extracting hierarchical relationships (like synonymy and hypernymy) from ordinary Japanese language dictionaries. By analyzing the structure and specific phrases within dictionary definitions, the authors aim to build a semantic thesaurus, a crucial component for advanced natural language processing.
What Problem Does This Paper Address?
Building a comprehensive thesaurus, which organizes words based on their semantic relationships, is vital for many natural language processing (NLP) tasks. However, manually creating such a resource is incredibly time-consuming and expensive. This paper tackles the challenge of automating this process by leveraging existing, readily available Japanese language dictionaries. Specifically, it focuses on extracting hierarchical relations like "is-a" (superordinate-hyponym) and "synonym" from the definition sentences (DSs) of these dictionaries. The core problem is how to reliably identify and classify these relationships from the often-unstructured text of dictionary entries.
Methods
The authors developed a pilot system that processes Japanese dictionary entries in several steps:
- Extraction of Entry Words (EW) and Definition Sentences (DS): The system first extracts the target word (EW) and its corresponding definition sentence (DS) from the dictionary. It also captures metadata like the part of speech and definition number.
- Normalization of DS: The extracted DSs are transformed into a more standardized Japanese sentence format to handle variations in punctuation and structure.
- Extraction of Functional Expressions (FE) and DW-EW Relation Information: The system identifies "Functional Expressions" (FEs) within the DS. FEs are specific phrases that often signal a semantic relationship between the EW and certain "Definition Words" (DWs) within the DS. A "Functional Word" (FW) is the core word within an FE. The system uses a "FW Dictionary" that maps FWs to their associated DW-EW relation types (e.g., hypernymy, synonymy) and the expected position of the DW.
- Extraction of Definition Words (DW) and DW-EW Relation Information: DWs are identified within the DS. The system uses a general word dictionary (containing about 75,000 noun words) and employs a longest matching method to find potential DWs. It also considers factors like whether the DW is modified and if there are multiple DWs.
- Decision of DW-EW Relation: Based on the identified DWs, FEs, and their associated information, the system decides the final DW-EW relation. If the relation is ambiguous, the DS is flagged for manual review.
The system categorizes DSs into two main types: TYPE I, where the final word is a DW, and TYPE II, where the final expression is an FE, with the DW preceding it. The paper also defines "Functional Patterns" (FPs) using regular expressions to help parse the structure of DSs and locate DWs and FEs.
Datasets
The primary dataset used in this study is the Shinmeikai Kokugojiten, a Japanese language dictionary published by Sansei-do, available in machine-readable form. This dictionary contains approximately 60,000 entries. The system was tested on 2,824 normalized DSs from this dictionary. A general word dictionary containing about 75,000 noun words was also utilized.
Key Insights and Ideas
- Dictionary Definitions as Rich Semantic Sources: The paper highlights that even ordinary dictionaries, beyond just providing definitions, contain structured clues (like specific phrases and word order) that can be mined for hierarchical semantic relationships.
- The Power of Functional Expressions (FEs): Identifying FEs and their associated Functional Words (FWs) is crucial. These FEs act as explicit markers for semantic relations, making automated extraction feasible.
- Two-Tiered Definition Structure: The classification of DSs into TYPE I and TYPE II provides a useful framework for systematically processing different definition styles and extracting relations.
- Challenges in Japanese Word Segmentation: The paper implicitly points out the difficulties in Japanese word segmentation, especially when dealing with hiragana, katakana, and unknown compound words, which require careful handling during DW extraction.
Key Results
The pilot system achieved promising results on the experimental dataset:
- High Extraction Ratio: Approximately 86% of the input DSs were successfully processed into output data.
- High Precision: The system achieved a precision of about 95% for correctly extracted DW-EW relations.
- Low Error Rate: The error ratio of incorrect output data was around 5%.
- Significant Check Data: About 14% of the input data was flagged as "check data," indicating cases where the system could not automatically determine the relation and required human intervention.
- Relation Distribution: The system successfully extracted various relations, including hypernymy (DW>EW), hyponymy (DW<EW), synonymy (DW=EW), and associative relations (R). The majority of extracted relations were hypernymy (1966 correct instances).
Limitations and Future Work
- Limitations:
- The paper acknowledges that most incorrect output data occurred during the extraction of DWs written in hiragana, due to limitations of the longest matching method.
- The system's performance is dependent on the quality and completeness of the dictionary itself. The authors note a "shortage of semantic information" in the dictionary, particularly regarding adequate DWs.
- Ambiguous relations require manual intervention.
- Future Work:
- Improve DW extraction, especially for hiragana notations.
- Reinforce the general word dictionary used for DW extraction.
- Further analysis of DSs to improve extraction accuracy.
- Development of systems for utilizing the dictionary more broadly, hierarchically structuring entry words, and constructing a practical-sized semantic dictionary with human assistance.
Summary updated