TL;DR
This paper surveys the techniques for building and applying Chinese knowledge graphs (KGs), highlighting their growing importance for AI and their potential impact on initiatives like China's "One Belt One Road" (OBOR). It details the stages of KG construction, from extraction to quality improvement and update, and showcases applications like semantic search and question answering, emphasizing the need for non-English KGs.
What Problem Does This Paper Address?
The paper addresses the growing need for robust Chinese knowledge graphs to support the advancement of artificial intelligence and its applications within China and in collaboration with countries involved in the "One Belt One Road" (OBOR) initiative. While English KGs are well-established, there's a significant gap in the development and application of KGs for other languages, particularly Chinese. This lack of resources hinders the development of intelligent applications and cross-cultural collaboration. The paper also highlights the challenge of integrating diverse data sources and ensuring the quality and currency of KGs, especially in a rapidly evolving information landscape.
Methods
The paper outlines a general framework for constructing Chinese knowledge graphs, comprising four key stages:
- Knowledge Extraction: This involves gathering data from structured (databases), semi-structured (XML), and unstructured (text) sources.
- Knowledge Integration: This stage focuses on linking and fusing knowledge from disparate sources. Techniques include:
- Knowledge Linking: Discovering semantic relations (e.g., equivalence, subclass) between entities and concepts. This includes monolingual linking (e.g., using EM algorithm for entity matching across Chinese encyclopedias) and cross-lingual linking (e.g., using linkage factor graphs and concept annotation to connect Chinese and English entities).
- Knowledge Fusion: Integrating equivalent knowledge by normalizing entities, attributes, and values that have the same meaning but different labels (e.g., unifying "birthday" and "birth date").
- Knowledge Quality Improvement: This crucial step addresses the inherent incompleteness and potential errors in KGs. It includes:
- Knowledge Completion: Inferring missing information, such as entity types (using attribute-driven type inference) and infobox facts (using hybrid LSTM-RNNs for text extraction).
- Knowledge Correction: Detecting and rectifying errors through rule-based methods and crowd-sourcing feedback.
- Knowledge Update: Mechanisms for keeping KGs current, including periodical updates and active updates that prioritize frequently changing or "hot" entities.
Datasets
The paper discusses the use of several large-scale Chinese encyclopedias as primary data sources for constructing knowledge graphs. These include:
- Baidu Baike
- Hudong Baike
- Chinese Wikipedia
Additionally, it mentions the use of data from social web sites for schema-level knowledge graphs and the integration of existing multilingual knowledge graphs like DBpedia and Yago. The paper also references the Linked Open Data Cloud as a significant global knowledge base.
Key Insights and Ideas
- The OBOR Initiative as a Catalyst: The paper frames the OBOR initiative as a significant driver for developing and applying Chinese KGs, fostering international cooperation in AI and knowledge sharing.
- Bridging the Language Divide: A core theme is the critical need for non-English KGs, with Chinese KGs serving as a valuable reference for other languages.
- Multilingualism in KG Construction: Techniques like cross-lingual linking and type inference are essential for building comprehensive KGs that span language barriers.
- The Importance of Data Quality: The paper strongly emphasizes that the utility of KGs hinges on robust methods for error detection, correction, and completion, especially given the crowdsourced nature of many data sources.
- Active Update for Dynamic Knowledge: The shift towards active update mechanisms highlights the challenge of keeping KGs current in a rapidly changing world, moving beyond simple periodic refreshes.
Key Results
The paper doesn't present specific quantitative results from a single experiment but rather surveys the state-of-the-art and the capabilities of existing Chinese KGs. It highlights:
- The existence of several large-scale Chinese encyclopedic KGs like Zhishi.me (over 10 million entities, 125 million triples) and CN-DBpedia (9.4 million entities, 80 million triples).
- The development of cross-lingual KGs like XLORE with over ten million bilingual entities.
- The successful application of these KGs in various intelligent applications, including:
- Semantic Search: Enhancing search engine results by understanding user intent (e.g., Baidu's search for "One Belt One Road").
- Question Answering: Providing direct and accurate answers to user queries (e.g., identifying the "initiative country" of OBOR).
- Intelligence Analysis and Decision Making: Facilitating fraud detection and other analytical tasks by integrating diverse data.
- Information Integration: Consolidating information from disparate sources into a unified view.
Limitations and Future Work
- Text-based Knowledge Extraction: The paper notes that few works effectively utilize Chinese open information extraction techniques from text, which is a crucial source for KG construction. This area warrants further study.
- Language-Independent vs. Dependent Approaches: While language-independent approaches are valuable, the paper acknowledges that some methods might benefit from incorporating Chinese linguistic characteristics.
- Knowledge Fusion Complexity: The paper mentions that current knowledge fusion strategies are relatively simple and can be prone to conflicts and errors, requiring human effort. More sophisticated language-independent approaches for knowledge fusion could be beneficial.
- Future Work: The authors suggest that the techniques and applications of Chinese KGs can serve as valuable references for developing non-English KGs in other languages, particularly for countries along the OBOR initiative. The OpenKG community's ongoing efforts to link KGs in vertical fields also represent a significant future direction.
Summary updated