Multi-mode Natural Language Processing for Extracting Open Knowledge

Abstract

As more and more open knowledge sources become available, it is interesting to explore opportunities of enhancing autonomous agents' capacities by utilizing the knowledge in these sources, instead of hand-coding knowledge for agents. A major challenge towards this goal lies in the translation of the open knowledge organized in multiple modes, unstructured or semi-structured, into the internal representations of agents. In this paper we present a set of multimode NLP techniques to formalize the open knowledge for autonomous agents. Two case studies are reported in which our robot KeJia, equipped with the multi-mode NLP techniques, succeeded in acquiring knowledge from the microwave oven manual and from the open knowledge database, OMICS, and solving problems that could not be solved before the robot acquired the knowledge.

A weighted causal theory for acquiring and utilizing open knowledge

“…Now we introduce the OMICS project and illustrate how to form an openknowledge base based on it. A complete specification on our service robot, KeJia [17,18] , to extract proper knowledge from OMICS, was provided in our previous work [11,12,19,20].…”

“…Note that, knowledge in OMICS are provided by semi-structured natural language sentences. We have provided an approach to converting elements in the Tasks/Steps table of OMICS to corresponding causal laws [11,12,19,20]. means a task or an action named τ i is accomplished or executed at time t i (1 ≤ i ≤ m), and t 1 ≤ · · · ≤ t m ≤ t. For instance, the element in Table 3 is converted to the following causal law: open(refrigerator) t 1 ∧ take(food) t 2 ⇒ get(food, refrigerator) t 3 , which means if the action open(refrigerator) is executed at time t 1 and the sub-task take(food) is accomplished at time t 2 , then the task get(food, refrigerator) would be accomplished at time t 3 (t 1 ≤ t 2 ≤ t 3 ).…”

“…In weighted causal theories, weights are directly assigned to candidate causal laws. There have been many efforts that share common concerns with open knowledge [8,9,10,11,12]. However, to authors' knowledge, this paper is the first work on formalizing the problem based on causal theories with weights.…”