CVA/Computational Team Fall 2001: CSE 676: preliminary coding of new words Marc Broklawski & Scott Napieralski: Analysis of Ehrlichs noun-definition algorithm English version of the N algorithm N alg (revised; Eng. Version) Report (if known): 1. Class membership:
a) In general, report most specific class of any class hierarchies Exception: always report basic-level class membership Possible exception: always report animal/plant maybe report: abstract/physical object b) Else report names of indivs that are Ns (continued) N algorithm (in English, revised), contd.
(Report:) 2. Properties If there are no known class memberships, then report props of indiv Ns as possible props of Ns 3. Structural, else possible structural, info I.e., possessive info
Should be: part/whole info (continued) N algorithm (in English, revised), contd (Report:) 4. Functions, else possible functions, of Ns I.e., things that Ns are used for 5. Acts, else possible acts, that Ns perform 6. Agents that perform acts on Ns & the acts they perform
7. Ownership info 8. Synonyms N alg: Comments Need to distinguish between: Gathering information Organizing & reporting information 2 different kinds of definitions: Human-like: report everything, even redundant info Lexicographic:
Constraints on how much info to report CSE 663 Accomplishments (Spring 2002) Marc Broklawski: Scott Napieralski:
Matt Sweeney: Rajeev Sood: Justin DelVecchio: Chris Garver: N algorithm; general resource `proximity `estuary V algorithm `proliferate `taciturn
Outline: 1. 2. 3. 4. What was your role in the project? What were your accomplishments? What are the immediate next steps? What is the longer-term future work? Marc Broklawski Role:
to make N algorithm more efficient; general resource for other students Accomplishments: made N algorithm more efficient case-frame dictionary for N algorithm recreated demos: cat, stender, brachet, tomato
instructions for how to run system Marc (continued) Next steps: global vars local vars `hackney demo (redo from scratch) Add demos to official demo list Solve innet/outnet problem
Marc (continued) Future work: SNePSwD? Better: Frans system explanation/debugging facility flag for 2 different kinds of definitions: complete/human-like lexicographic Scott Napieralski: `proximity 1. 2.
3. Galileo recently flew just 120 miles (200 kilometers) above Europa, a proximity which allowed the spacecraft to take the most detailed pictures ever of the mysterious satellite. Archeologists who discovered and excavated the small cave in 1983 had assumed that the brittle black material on the skulls was asphalt because of its color and the site's proximity to the largest asphalt deposits in Israel. Because of its proximity, Earth's only natural satellite the Moonbecame the first celestial body to be visited
by humans (in 1969). Scott (continued) Accomplishments: Re-represented 3 passages for Ehrlichs N alg Ran alg to look for improvements Discovered that alg doesnt find everything it could Scott (continued) Next step: Modify N alg to report more info
Future work: Look for more improvements in N alg Look for common traits in report info Assume N is what X, Y, & Z are; X, Y, & Z have props P, Q, R; Therefore, Ns have P, Q, R Need more background knowledge Abstract Ns: No structure, actions, etc., which are heart of N alg Therefore, need other categories for abstract Ns Matt Sweeney: `estuary
An estuary is a coastal area where fresh water from rivers and streams mixes with salt water from the ocean. Many bays, sounds, and lagoons are estuaries. Estuaries provide safe spawning grounds and nurseries and are critical for fish, birds, and other wildlife. Matt (continued) Role: Get practical (rather than good) representation to work Change code so that code and input would
work well together Matt (continued) Accomplishments: Coded `estuary passage & background knowledge into SNePSUL Modified N algorithm to get necessary-condition definition Discovered necessity of tricking forward inference to work Matt (continued) Next steps:
Create sufficient-condition definition function Test with SNePS 2.6 Develop tracing algorithm For explanation facility Good Lisp exercise Modify N algorithm to report more detail E.g., about agent-act-object case frame Possibly solvable by NLG facility Need def to include info about bays, sounds, lagoons Matt (continued)
Future work: Optimal representation, maybe in SNePS 3 Continue work with tracing algorithm Rajeev Sood: Verb algorithm Role: Understand V algorithm Translate into English For use by education team Create case-frame dictionary for V algorithm Rajeev (continued)
Accomplishments: Met with Justin Analyzed V algorithm Improved V algorithm a bit Created English translation Created case-frame dictionary for V algorithm
English translation of V alg: 1. 2. 2. 1. Categorize the subject and report it. Determine the form of the verb: a. If bitransitive, categorize the direct and indirect objects and report them. b. If transitive, categorize the direct object and report it. c.
If reflexive, report that the subject performs that act upon itself. d. If intransitive, go to the next step. 3. 3. Find anything that the verb occurring enables (cause function) and report it. 4. 4. Find anything that occurs that enables the verb to occur Revised Eng. Transln of V alg:
1. Categorize the subject and report it. 2. If it exists, categorize the direct object and report it. 3. If it exists, categorize the indirect object and report it. 4. Find anything that the verb occurring enables (cause function) and report it. 5. Find anything that occurs that enables the verb to occur (effect function) and report it. Rajeev (continued) Next steps:
Change setqs to lets Tracing function for V alg Need more info returned How to determine V category from structure Rajeev (continued) Future work: V alg needs LOTS of elaboration!!! More than just cause/effect info needed
E.g., throw = ball leaves hand, goes to other hand, etc. Parser/generator Justin DelVecchio: `proliferate A decade ago research on lab animals revealed that stem cells taken from animal embryos are astoundingly versatile. They grow in the lab, proliferate like rabbits and turn into specialized cells such as neurons.
Justin (continued) Role A verb! 4 goals: Redo Marcs representation for Ehrlich case frames Work with Rajeev Sood to analyze V algorithm Background knowledge Find more passages
Justin (continued) Accomplishments: Studied what Marc had done Met with Rajeev to understand V algorithm Created `throw demo to show how V algorithm works Re-represented `proliferate passage to match V alg
Coded background knowledge Esp. actions common to stem cells & rabbits Tested V algorithm on passage + background kn. Analysis of changes in V algorithm Justin (continued) Next steps: More work on background knowledge: More actions & characteristics of stem cells & rabbits Changes to V algorithm: Better output (maybe in English)
Delete labeling of Vs (transitive, etc.); replace w/ node structure Differentiate results for intransitive & transitive Vs like `throw Justin (continued) Future work: Need more Vs!
More passages with `proliferate More/better background knowledge Parser/generator! Chris Garver: `taciturn Unlike his brothers, who were noisy, outgoing, and very talkative, Fred was quite taciturn. Chris (continued) Role: Adjective!
Accomplishments: Re-represented passage in SNePSUL Represented background knowledge in SNePSUL Looked at other passages & their background knowledge Outline of Adj algorithm Some new case frames Outline for Adj(/Adv?) Alg: 1. Synonyms & antonyms Problem: How to determine!
2. Class membership E.g., hot isa temperature; quick isa speed 3. Lists of objs known to be describable & not describable w/ the Adj 4. Lists of actions that can/cant be performed while possessing the Adj
someone is not communicative & talks very little This example raises issue of negative adjs. Chris (continued) Next steps: Fix bugs in implementation Esp. SNeBR problem: unexpected contradiction Need tracing feature! Need positive passages Chris (continued) Future work:
Analysis of adjs: Categories/possible contexts/domains of adjs Kinds of Ns that certain adjs describe What does adj refer to: it was heated Implement Adj algorithm Possibly combine with Adv algorithm Rep & test more passages Summer Plans: Scott: Re-implement N alg
with explanation/debugging facility with good documentation! Justin: Re-implement V alg Use Levins theory of V categories? Work on English translation Rapaport & Kibby: New grant proposal
Orlando conference; Scotland conference(?) Literature-review paper? Future Work: Need parser/generator! Implement Frans automated belief revision system! Implement all examples from the CVA literature Use OpenCYC to supplement background knowledge?
