Intelligent reasoning on natural language data: A non-axiomatic reasoning system approach.

摘要

Research on Artificial General Intelligence has re-gained attention since the 2000s with a range of feedback from other disciplines, such as neurology, cognitive science, linguistics, psychology, philosophy, and such. NARS, a non-axiomatic reasoning system, is a general-purpose intelligent system able to work with insufficient knowledge and resources, and to adapt to its environment by learning from experience. It treats intelligence as a domain-independent capability with no domain-specific sub-module. Since the human mind evolved under the same restriction, this normative model displays many human-like properties.;NARS is used to reinterpret several well-known results in cognitive science, such as Wason's selection task, the Linda problem, and U-shaped learning, which cannot be explained by traditional normative models, but can now be handled by NARS in a unified way. This study specifically investigates the reasoning capabilities of NARS, a non-axiomatic reasoning system, on natural language data. NARS is used to mimic U-shaped learning of passive voice in English, subjective pronoun resolution, and contextual dependency of concepts. For this purpose, logical form from WordNet is translated to NARS. Furthermore, a convolutional neural network, which is available online and trained with images from ImageNet, is used to recognize possible noun categories of a given image.;The results have shown that a general-purpose system can simulate human-level behavior on language data without a built-in linguistic module.