Semantic category acquisition in dialogue for interactive object learning

Abstract

An important aspect of humanoid robots in a natural environment is the ability to acquire new knowledge through learning mechanisms, which enhances an artificial system with the ability to adapt to a changing or new environment. In contrast to most learning algorithms applied in machine learning today, which mainly work with offline learning on training samples, such learning mechanisms need to be performed autonomously and through interaction with the environment or with other agents/humans. In this paper we describe a learning algorithm as a dialogue approach for learning semantic categories and object description in object learning. New objects are introduced to the robot and learning dialogues are conducted as a means of information acquisition. In dialogue, the robot can acquire semantic categories, type and properties of objects, learn new words for object descriptions and learn and association to visual identification from object recognition. In contrast to existing work, this approach combines recognition of real objects, new words learning and semantic categories in one learning dialogue. The presented approach has been implemented in a dialogue system and evaluated on the humanoid robot Armar III.