On Solomon Its Context blue sky poc Selmer
On Solomon & Its Context (blue sky p-o-c) Selmer Bringsjord Micah Clark & the RAIR Lab braintrust Rensselaer AI & Reasoning (RAIR) Laboratory Department of Cognitive Science Department of Computer Science Rensselaer Polytechnic Institute (RPI) Troy NY 12180 US Santa Fe 11. 1. 06
The RAIR Lab. . .
http: //www. cogsci. rpi. edu/research/rair
• • RAIR Lab Method Isolate and dissect human ingenuity. • • Formalize weak correlate to this ingenuity in advanced logical systems. • • • Psychology of Reasoning is central. Info professionals, yes, but IAs are pro reasoners paid to produce determinate output. Subsumes elementary extensional logics (description logics, standard FOL, logic programming/Prolog); more expressive extensional logics (SOL); defeasible extensional logic (e. g. , non-monotonic logic); intensional logics needed formalizing concepts like betrayal, revenge, deception; probabilistic/strength-factor-based reasoning and inductive logic; and visual logics. Implement correlate in working computer programs. Augment correlate as needed with machinespecific power.
Some Implications • • QA for us (a brand new foray) must take place in the context of logical systems. Specifically, text or diagrams/pictures must be expressed as formulas and associated structures (e. g. , proofs, arguments, models) in logical systems.
Visual Knowledge in Wargaming/Prediction (Psy. Pre)
Prior DTO-Sponsored R&D: Slate 2. 0 (NIMD/ARIVA)
Slate: Some Functionality
• • Reasoning with Uncertainty We do not take a probabilistic approach, but rather a more cognitively plausible strength factor-based approach. This is more in line with what real humans normally do in real life. • • • The empirical evidence against the view that humans use probabilities is overwhelming. Notice that strength factors underlie CIA’s Tradecraft Review. We turn to Roderick Chisholm’s Theory of Knowledge, formalize his scheme, and implement the formalization in Slate.
Chisholm’s Strength Levels Certain Evident Beyond Reasonable Doubt Probable Counterbalanced Probably false Reasonable to disbelieve Evidentally false Certainly false
“Philly Bomb” wmv. . .
IKRIS (interoperability)
Basic Interoperability Challenge 2 “Common Logic, ” an DTO standard Translators IKL automatic translation 3 4 1 Chart F from Slate is sent out toward SNARK. Some hypothesis H can now be confirmed by SNARK. KB KB
KANI Noöscape Slate Using KANI, Analyst considers proposition : there is some meeting A query is sent from KANI to Noöscape asking for hypotheses. Noöscape performs abduction on the input combined with , and generates the hypothesis that Pakes cannot leave the ship. is returned to KANI Analyst considers hypothesis : Pakes can’t attend the meeting A query is sent from KANI to Slate asking for whether and entail. Slate does not confirm the hypothesis, but is able to generate a counter-model such that. is returned to KANI Analyst considers the possibility that does hold. A query is sent from KANI to Slate asking for hypotheses using and. • • • Slate and Noöscape find a hypothesis in which the references to the ‘meeting’ actually refer to a terrorist attack. is returned to KANI Analyst says, “Aha!” IKRIS Capstone Dance - v 2
For ARIVA: AKRRIV Quartet 1. Engineer specs and implementation allowing interoperability between ARIVA systems (and, for that matter, DTO and IC systems, period). • Key: Vivid-CL 2. Engineer specs and implementation allowing an analyst to be modeled by ARIVA systems (and, for that matter, DTO and IC systems, period), so that such systems can enter into genuine dialogue with analysts. • Key: RASCALSIA 3. Engineer specs and implementation for an “operating system” to allow interconnected ARIVA systems (and, for that matter. . . ) • Key: Director 4. Prove that these specs and this engineering works with and enhances Slate, and ditto for other ARIVA & DTO/IC systems.
Solomon: A Next-Gen Q&A System Rensselaer Polytechnic Institute (Dr. Selmer Bringsjord – PI) Reasoning Human Defensible Natural Knowledge Suppositional – Computer Answers, over Acquisition Visual Collaboration Rational Reasoning & via. Symbolic Reading Justifications, via Knowledge Visual & Symbolic Conversation & Intuitive. Read Explanations Read Information Solomon Can… Can. . . ? ? Possibly, if…if… Suppose that… No, because… Produce Proofs, Linguistic Non-monotonic Textual Visual Information Arguments, / Symbolic Arguments, & Introduction Representation of of & English Information Reports Hypotheticals Conversational Discourse PLANS We leverage of our adv. IA assistant Slate (DTO) • Extend Slate with our Learning by Reading (LBR) technology (DARPA IPTO) and our Vivid family of visual logics (DARPA IPTO) • Formalize a non-monotonic discourse representation theory for conversational Q&A • Develop a conversation-driven UI Agent: PEO STRI COTR: Agency/Name GOALS Construct p-o-c Q&A system with the attributes: • Knowledge Acquisition via Reading • Human – Computer Collaboration via Conversation • Natural Suppositional Reasoning • Defensible Answers, Rational Justifications, & Intuitive Explanations • Reasoning over Visual & Symbolic Knowledge MILESTONES A sequence of six demonstrations: 01 -01 -07 Knowledge Acquisition from Text 03 -01 -07 Knowledge Acquisition from Visual Data 05 -01 -07 NLG of Justifications and Explanations 07 -15 -07 Interactive Conversation (v 1) 09 -02 -07 Interactive Conversation (v 2) 10 -31 -07 Capstone end-to-end Demonstration October 2006
Machine Reading of Text. . .
Machine Reading
Machine Reading of Visual Data. . . (wmv)
- Slides: 24