TL;DR
This paper introduces a novel Korean knowledge extraction system designed to enrich KBox, a knowledge base built upon Korean DBpedia. The system offers a user-friendly web interface, enabling anytime, anywhere knowledge extraction and accumulation, thereby laying the groundwork for an iterative knowledge learning system.
What Problem Does This Paper Address?
The increasing demand for structured knowledge has fueled interest in information extraction (IE) from natural language. While large-scale knowledge bases (KBs) like Freebase and DBpedia exist, they often lack comprehensive knowledge, necessitating knowledge completion tasks. A significant challenge is that knowledge is often language-specific; for instance, Korean Wikipedia holds unique cultural information not readily available in English Wikipedia. Furthermore, building KBs for specific languages requires tailored ontology schemas and extraction systems. This paper addresses the gap by proposing a Korean-specific knowledge extraction system that can continuously enrich a Korean KB (KBox) through an accessible web interface.
Methods
The proposed system comprises three main parts: Preprocessing, Relation Extraction (RE), and Post-processing.
- Preprocessing: This stage involves using an NLP Tool to extract features like part-of-speech tags and dependency parses. A key component is Entity Linking, which uses a bidirectional LSTM for entity boundary detection and an SVM for disambiguation, linking text mentions to existing entities in KBox. Notably, single-character Korean entities, which can have multiple meanings, are not treated as candidate mentions to avoid ambiguity. An Input Formatter then prepares data for the RE models.
- Relation Extraction (RE): Recognizing that a single RE module often yields unsatisfactory performance, the system employs an ensemble of multiple RE models. These include:
- Pattern-based RE: Aims for high reliability by using human-annotated patterns, but suffers from low recall.
- Sentence-level RE: Utilizes CNN and LSTM models with distant supervision to improve recall and scalability.
-
Paragraph-level RE: Addresses Korean's zero anaphora (omitted subjects) by incorporating global contexts from paragraph-level embeddings and using graph analysis to identify central entities. The system also considers Named Entity Recognition (NER) results for entities not yet present in KBox, focusing on Person, Location, and Organization types as potential new entities.
-
Post-processing: This stage combines the outputs from the various RE modules. An Ensemble module prioritizes knowledge with high scores from individual modules or knowledge extracted by multiple modules. Schema Filtering then removes invalid triples that do not conform to the ontology's domain and range definitions. Finally, an Output Formatter generates JSON for the web interface and tab-separated values for KBox integration.
Datasets
The paper mentions using the Korean DBpedia and Wikidata for initial entity definition and mapping local properties to ontological ones. It also refers to the use of distant supervision for training RE models, which leverages existing large-scale KBs to automatically label sentences. No new benchmark datasets are explicitly introduced.
Key Insights and Ideas
- Language-Specific Knowledge Bases are Crucial: The paper highlights that knowledge is often language-bound, necessitating dedicated systems for languages like Korean to capture unique cultural and linguistic nuances.
- Ensemble Approach for Robust RE: Relying on a single RE module is insufficient. Combining multiple models with different strengths (e.g., precision vs. recall, sentence vs. paragraph level) leads to more comprehensive knowledge extraction.
- Handling Korean Linguistic Peculiarities: The system's approach to single-character entities and the paragraph-level RE model for zero anaphora demonstrate a thoughtful consideration of Korean language characteristics.
- Web Interface for Accessibility and Iteration: The web interface is a key innovation, democratizing knowledge extraction and enabling continuous accumulation, which is vital for iterative learning and system improvement.
Key Results
The paper presents a demo system and its architecture. While specific quantitative results (e.g., F-scores for the ensemble) are not detailed, the system's contributions are highlighted:
- The first open Korean knowledge extraction system with a web interface.
- Immediate accumulation of extracted knowledge into KBox.
- The system is released under a CC BY-NC-SA license.
Limitations and Future Work
- Limitations: The paper acknowledges that even state-of-the-art RE models show low performance (F-scores 40%-50%). The pattern-based RE model has low recall, and the distant supervision approach for sentence-level RE can introduce noise. The entity linking system does not treat single-character entities as candidates due to ambiguity.
- Future Work: The ultimate goal is to build an iterative knowledge learning and extraction system. The continuously accumulated knowledge in KBox is intended to help the core knowledge extraction modules (entity linking and RE) learn and improve over time.
Summary updated