Emergence of analogy from relation learning

Area of Science:

• Cognitive Science
• Computational Linguistics
• Artificial Intelligence

Background:

• Humans develop abstract semantic relation understanding by middle childhood.
• A key challenge is explaining how abstract relations emerge from non-relational inputs for symbolic reasoning.
• Existing models struggle with the feature alignment problem in abstract relation representation.

Purpose of the Study:

• To develop a computational model that learns abstract semantic relations from lexical data.
• To demonstrate how deep learning and supervised learning can synergize for relation extraction.
• To enable relational generalization and analogical reasoning from non-relational inputs.

Main Methods:

• A computational model combining deep learning ('big data') for word features and supervised learning ('small data') for relation representations.
• Input: Labeled word pairs from deep learning; Model creates augmented representations by remapping features based on value differences.
• Weight distributions are extracted to estimate relation probabilities and capture human typicality judgments.

Main Results:

• The model effectively addresses the feature alignment problem for diverse abstract semantic relations.
• It accurately captures human typicality judgments for various semantic relations.
• Relational similarity measure enables solving verbal analogies with human-level accuracy.

Conclusions:

• Abstract semantic relations can be induced by bootstrapping from non-relational inputs.
• The model's modular representation supports basic symbolic operations, like forming converse relations.
• This approach provides a pathway towards protosymbolic representation systems and robust analogical reasoning.