Knowledge extraction: a comparison between symbolic and connectionist methods

Area of Science:

• Artificial Intelligence
• Machine Learning
• Knowledge Representation

Background:

• Effective knowledge representation is crucial for user understanding and adoption of AI systems.
• Both symbolic and connectionist AI approaches offer techniques for knowledge extraction.
• Existing methods vary in their ability to translate learned information into accessible formats.

Purpose of the Study:

• To investigate and compare three distinct knowledge extraction techniques.
• To evaluate methods based on symbolic learning and neural network knowledge extraction.
• To assess the suitability of different techniques for generating linguistic knowledge representations.

Main Methods:

• The study examines the C4.5 algorithm for decision tree and rule induction.
• It analyzes the CN2 algorithm for inducing if...then rules from datasets.
• The TREPAN algorithm is investigated for extracting decision trees from trained neural networks.

Main Results:

• C4.5 and CN2 extract knowledge directly from datasets, producing rules or decision trees.
• TREPAN extracts knowledge in the form of decision trees from pre-trained neural networks.
• All three methods utilize decision trees or rules as their knowledge representation.

Conclusions:

• The C4.5, CN2, and TREPAN algorithms offer diverse approaches to knowledge extraction in AI.
• Understanding these techniques is key to developing AI systems that are transparent and user-friendly.
• The choice of technique impacts how learned knowledge is represented and communicated.