Extracting Word Relationships from Unstructured Data
Robots are advancing rapidly in their behavioural functionality allowing them to perform sophisticated tasks. However, their ability to take Natural Language instructions is still in its infancy. Parsing, Semantic Intrepretation and Dialogue Management are typically performed only on a limited set of primitives, thus limiting the set of instructions that could be given to a robot. This limits a robot’s applicability in unconstrained natural environments (like households and offices) [8]. In this project, we are only addressing the problem of semantic interpretation of human instructions. Specifically, our Extracto algorithm provides a method to extract potential actions (verbs) that could be performed given two household objects (nouns). For example, given the nouns “Coffee” and “Cup”, Extracto identifies the action (verb) “pour” indicating that ‘coffee should be poured in a cup’, and not ‘stored’ or ‘roasted’. A human instruction “I want coffee” or “Get me a cup of coffee” is a goal for the robot, but does not specifically instruct the robot what to do with the cup and the coffee. The Extracto method helps address this particular problem. In addition, given an action (verb), Extracto identifies the most suitable objects (nouns) to perform the task. For example, given an action (verb) “introduce”, Extracto identifies a series of suitable noun pairs, one of which is “friend” and “host” which means ‘a friend is to be intro0 a [email protected] 0 b [email protected] 0 c [email protected] duced by a host’
Research Paper Link: Download Paper