Simple Language Learning in Artificial General Intelligence

Abstract

The development of Artificial Intelligence (AI) has made it possible for computers to solve more and more of the tasks that humans can perform. One area of interest within AI research is Artificial General Intelligence (AGI), which aims to develop AI systems capable of solving multiple varying tasks. The Generic Animat research project is one biologically inspired approach to AGI. Our thesis aims to explore Natural Language Processing within the Animat model, by mimicking the way young children start to learn their first language. During the course of the thesis project we have developed two slightly different prototypes (referred to as the “Temporal Animat” and the “Chunking Animat”), both of which can learn to recognise and produce words that they are exposed to in a noise-free environment. Furthermore, functionality for word-to-word association and multimodal association has been implemented, thereby allowing the prototypes to associate between words, based on their occurences in some input text, and between words and other types of sensor input representing non-language concepts in the Animat’s environment. The prototypes were evaluated on small input texts with a limited number of words. The result of these evaluations suggests that both versions of the Animat are capable of achieving high degrees of accuracy when their word-to-word associations are compared to those of vector space models. The evaluation of multimodal associations has been done only for the Temporal Animat. The results indicate that the Temporal Animat makes mostly accurate associations between words and other input, although more tests on larger data sets are needed to determine how well it works on a larger scale. Due to the similarity in implementation between the two prototypes, the Chunking Animat is expected to be equally successful at this task.