Collaborating NLP Agents

Authors

• Miki Mizutani
• Leonidas Gee

Abstract

The project involves the design and implementation of a multi-agent system for training classifiers with different configurations given a parameter space. Using a master-slave architecture, the system searches the parameter space by random sampling and trains the classifiers for named-entity recognition on the CoNLL-2003 dataset in a round-robin manner. The best trained classifier is then determined and evaluated on the test set. Results show that the proposed multi-agent system is able to search the parameter space efficiently.

Outcome

• project forum
• repository
• final report