Learning Common Sense Knowledge Through Human-Robot Dialogue
Current deep learning methods perform exceptionally well for pattern recognition and synthesis tasks in areas such as speech processing and computer vision. Artificial Neural Networks (ANN) can easily train on high dimensional real-world data that is often noisy in nature. However, there are several limitations related to current approaches. One of the shortcomings is the black-box or the explainability problem. Due to characteristic lack of human-readable representation of deep learning architectures, it is difficult to extract an explanation for an inference made by a system. Another significant issue, that is particularly of interest for robotics, is the limited capability of relational reasoning. These problems contribute to limited applicability of neural network based approaches for Natural Language Processing (NLP) applications.
The video demonstrates a hybrid framework that combines ANN models with logic-based learning for common-sense reasoning. Inductive Logic Programming (ILP) involves learning of a hypothesis that, together with background knowledge, explains a set of positive and negative examples. The learned hypothesis (when added to the background knowledge) can then be used for inference.
The framework uses NN architecture to obtain speech string from audio input and extract a dependency tree through natural language parsing. Using the dependency tree, a structured representation is obtained for reasoning. The system is then able to infer an answer for user question using its knowledge base. In case the answer was wrong, it can learn a hypothesis using factual knowledge based on experience and observation.