Exploring the Utility of ResearchCyc for Reasoning from Natural Language
Overview
This project is working on focussed studies showing how much value can
be achieved from use of ResearchCyc for semantic tasks in Natural
Language Processing.
Progress
- Oct 2004
- We've got and installed ResearchCyc and are starting to learn how to
use it.
- Dec 2005
- We decided to focus first on exploring the utility of ResearchCyc to help in
performing local textual inferences, of a sort that we are investigating
in the context of the PASCAL Recognizing Textual Entailment (RTE) challenge.
- Jan 2005
- We're having some troubles getting CycNL working to do a text to
knowledge representation mapping.
- May 2005
- We can now explore Cyc-space along genls and isa relations
to inform our judgement of semantic similarity among verbs. We're investigating how robust and useful this is in the context of RTE. We're starting use more of ResearchCyc's information as a measure of similarity via general lexical relationships:
this anticipates the likely scenario that ResearchCyc can't parse the target entailment sentences, and we can't just ask for a logical proof of entailment.
- Jul 2005
- We've prepared some error analysis on several entailments our current (non-ResearchCyc)
system misses and that seem relatively straightforward and representative. The analysis is meant to understand what
ResearchCyc might usefully provide or what's crucially missing.
- Sep 2005
- We've completed some experiments on the utility of
ResearchCyc-derived similarity measures in textual inference type
problems.
Presentations
Chris Cox's Cyc presentation, given at NLP lunch,
explains some Cyc basics and how to get connected to ResearchCyc at Stanford. Several graphics and examples
are taken from the more extensive ResearchCyc tutorial here.
Reports
-
Christopher Cox. 2005. Assessing the
Utility of ResearchCyc in Recognizing Textual Entailment.
[html]
[doc]
[pdf]
Contact Chris Cox
(or Chris Manning or
Andrew Ng) for further information.