Addressing Label Sparsity With Class-Level Common Sense for Google Maps.

Area of Science:

• Artificial Intelligence
• Knowledge Representation and Reasoning
• Machine Learning

Background:

• Traditional knowledge graph (KG) curation relies on logic-based methods, neglecting the common-sense associations crucial for practical applications.
• The machine learning community faces a significant challenge with label sparsity, hindering AI development in data-scarce domains.
• Acquiring knowledge and data for AI remains a critical bottleneck, echoing challenges from the expert systems era.

Purpose of the Study:

• To propose a novel, crowd-sourced approach for acquiring class-level attributes to address the label sparsity problem.
• To enhance knowledge graphs with common-sense associations between categories.
• To enable machine learning systems to function effectively in low-data environments.

Main Methods:

• A three-tier crowd-based approach for acquiring class-level attributes representing common-sense associations.
• Utilizing the classic knowledge-base default and override technique for knowledge acquisition.
• Demonstrating the approach on industrial-scale problems involving augmenting KGs of places and offerings.

Main Results:

• Successfully acquired class-level attributes, effectively mitigating the label sparsity problem.
• Augmented existing knowledge graphs with crucial associations between places and offerings (e.g., products in stores, dishes in restaurants).
• The developed approach was instrumental in enabling a worldwide local search capability on Google Maps.

Conclusions:

• The proposed three-tier crowd approach offers a scalable and effective solution for knowledge acquisition, particularly for class-level attributes.
• Addressing label sparsity through common-sense associations is vital for advancing AI and machine learning applications.
• This methodology has demonstrated real-world impact by enhancing local search functionalities and improving user access to information.