Knowledge Representation and Bayesian Inference for Response to Situations

Contents

Commonsense reasoning is crucial in making humanoid robots capable of responding to situations in a human-like fashion. To address this challenge, we have used a Bayesian Network to compare different responses to find a likely response. This Bayesian Network is populated for the situation under consideration from a multidimensional semantic net, called the PraxiNet. PraxiNet is used to graphically represent all possible situations and responses. Instead of manually engineering the knowledge base for PraxiNet, we have used distributed knowledge capture techniques as the knowledge source for PraxiNet. We collect knowledge from volunteers over the web about causality and responses to situations. This knowledge is very noisy and is processed using Natural Language Processing (NLP) techniques including spell checking, heuristic-based pattern removal and chunking to improve the quality of knowledge. PraxiNet is expanded using WordNet and a thesaurus, and subsequently condensed by lemmatization, synonym and hypernym merging to increase the overlap of knowledge and the density of the network. Given a situation (or multiple situations) we extract the relevant part of PraxiNet into the Bayesian Network for computation of suitable responses. This approach is scalable and can handle millions of pieces of knowledge to find the common sense responses for a given situation.