Incorporating World Knowledge to Document Clustering via Heterogeneous Information Networks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Natural Language Processing

Background:

• Supervising machine learning protocols is costly, often requiring domain experts.
• World knowledge, or general-purpose knowledge, offers a potential solution for indirect supervision.
• Adapting and representing world knowledge for specific domains are key challenges.

Purpose of the Study:

• To explore the use of world knowledge for domain-dependent document clustering.
• To develop methods for adapting and representing world knowledge in machine learning.
• To propose a novel clustering algorithm incorporating world knowledge.

Main Methods:

• Utilized Freebase and YAGO2 as sources of world knowledge.
• Developed three methods to specify world knowledge for domain adaptation, resolving entity and type ambiguity.
• Represented data using heterogeneous information networks incorporating world knowledge.
• Proposed a clustering algorithm capable of handling multiple data types and subtype constraints.

Main Results:

• Experimental results on 20newsgroups and RCV1 datasets demonstrated significant performance improvements.
• Incorporating world knowledge as indirect supervision outperformed state-of-the-art clustering algorithms.
• The proposed method surpassed traditional clustering algorithms enhanced with world knowledge features.

Conclusions:

• World knowledge can effectively serve as indirect supervision in machine learning tasks.
• The proposed framework offers a viable and efficient alternative to expert-driven supervision for document clustering.
• This research opens avenues for leveraging external knowledge bases to enhance domain-specific learning protocols.