From Disambiguation Failures to Common-Sense Knowledge Acquisition: A Day in the Life of an Ontological Semantic System

The paper describes a semi-automatic method of identifying common-sense knowledge by running an ontological semantic system and focusing on its failures to interpret sentences that a human is not challenged by. Without common-sense knowledge, "He put a banana in his trunk" produces 6 representations, corresponding to 6 senses of "trunk" recorded in the lexicon and anchored in different ontological concepts - roughly, car-part, elephant-part, tree-part, torso, luggage, software-term. A human language user reduces them to 3, using several pieces of common-sense knowledge that should therefore be added to the system resources to improve its performance. The significant result is that by running the system not only experimentally but also in real-life applications, we ensure an ongoing test against the lack of common-sense knowledge, at least some of which is known to be captured by the system, if it interprets the text correctly and not captured when the interpretation is inadequate.

I. Introduction

This paper describes a hybrid human-computer effort to collect common-sense knowledge from natural language text taken from the web. The majority of information on the internet contains natural language text and is not as useful as it could be if a computer were to actually understand what the texts are about and, for instance, to assert salient information from multiple documents.