Computational Contextual Vocabulary Acquisition: A Noun Algorithm
Contents
1. J Rapaport Scott T Napieralski and I derived a streamlined English version of noun algorithm Scott followed up this streamlined version with a more detailed English version as well 2 VARIOUS DETAILED INSTRUCTIONS This section includes important running instructions for the verb and noun algorithms and SNePS on a UB CSE department machine and detailed instructions on creating demo files for passages 2 1 Instructions for Running CVA Algorithm and SNePS on UB CSE Machines e Start Lisp o From within Emacs or XEmacs press M x the M x is generated by pressing the Esc x keys then type run old acl Note This will ONLY work if you have already added load cselisp el to your emacs file click here for a sample emacs file o From shell prompt type old acl e To load SNePS at the Lisp prompt type o load projects snwiz bin sneps or o ld projects snwiz bin sneps e At the Lisp prompt to load the noun algorithm type o load projects rapaport CVA mkb3 CVA src fast code or o iid projects rapaport CV A mkb3 CVA src fast code e Now start SNePS by typing sneps at the Lisp prompt e Load your demo CVA passage by typing demo lt pathname gt lt filename gt at the SNePS prompt o lt pathname gt path to directory where demo file is found o lt filename gt demo file name o Example CVA demos can be found in proj ects rapaport C VA mkb3 CV A demos Refer to these demos to see what must be2. We have thus come to the conclusion that the only way to proceed from here is to begin to recreate the hackney demo by ourselves This should be complete in the near future Finally it is also highly suggested that the outnet and innet functions of SNePS not be used For example since at the beginning of the brachet demo run all previous network structures built in the cat demo should be present including background information it should be the case that at the end of the cat demo the existing network should be outnetted to a file So if we wanted to run the brachet demo a few days later then we would not have to run the entire cat demo again before running the brachet demo We could simply innet the network that was outnetted during the completion of the cat demo run and begin running the brachet demo However this seems to somehow significantly degrade the speed at which any type of deductions are done when reading through any subsequent demos in fact sometimes a deduction will completely hang up the system So the moral of the story for now is that you will need to run all previous 24 demo runs that a current demo is dependent on in Ehrlich s demos However it has currently been discussed during research meetings with Professor Rapaport that it may not be necessary for Ehrlich s demos to be dependent on previous runs to function correctly give the right results Note that this is pure speculation and a further investigati
3. Computational Contextual Vocabulary Acquisition A Noun Algorithm Marc K Broklawski State University of New York at Buffalo Department of Computer Science and Engineering CSE 663 Advanced Knowledge Representation May 3 2002 ABSTRACT The purpose of my work throughout the current academic year related to making Karen Ehrlich s noun algorithm more efficient A great amount of effort has been made to remove all the redundancy present in the original noun algorithm Additionally several miscellaneous tasks have been completed 1 detailed instructions for running SNePS on our local CSE machines i e all graduate machines 2 a dictionary of case frames for the noun algorithm those that the algorithm actually looks for 3 detailed instructions for actually creating a demo to be run with Ehrlich s verb and noun algorithm 4 a illustrated version of the noun algorithm and 5 recreation of several of Ehrlich s original demos Finally I discuss future tasks related to both changes I have made and ones that need to be made in the future 1 INTRODUCTION Ongoing research involving the development of a CVA curricula strategy for middle school and high school students is being investigated by Rapaport and Kibby 2001 which focuses on SMET texts This research is looking to extend and enhance an algorithm that figures out the meaning of an unknown word nouns and verbs for now from context developed by Ehrlich 1995 in which
4. ct build object Joe property build lex smart Additional Modal Concepts Possibly Joe is smart Sir Joe must bring the hart to the hall Sir Joe shall bring the hart to the hall All knights who wished to joust might joust Sir Joe can enter the lodge Sir Joe should not joust against Sir Marc Merlin said that King Arthur ought to slay his brachet qpossibly gt sometimes automatically added when doing belief revision e g smite demo 14 3 7 object proper name object proper name Syntax assert object i proper name j Semantics m is the proposition that i has the proper name j Sample Context Jack is heavy describe assert object Jack proper name build lex Jack describe assert object Jack property build lex heavy 15 3 8 object property object property Syntax assert object 1 property j Semantics m is the proposition that i has the property j Sample Context Jack is heavy describe assert object Jack proper name build lex Jack describe assert object Jack property build lex heavy 16 3 9 object rel possessor object rel possessor Syntax assert object i rel j possessor k Semantics m is the proposition that i is k s D Sample Context Pyewacket is Evelyn s cat describe assert object Pye proper name build lex Pyewacket describe assert object Eve proper name build lex E
5. d of using ancient ones In addition the outnet and innet problem described in Section 5 should be discussed and shown to Prof Stuart Shapiro Hopefully this problem can be resolved 28 and make the creation of demos easier Also a meeting should be made with the current SNWIZ Fran Johnson as of now to add the current demos in my projects rapaport C VA mkb3 CVA demos folder to the default SNePS demos REFERENCES 1 Ehrlich K 1995 Automatic Vocabulary Expansion through Narrative Context TR 95 09 Buffalo SUNY Buffalo Dept of Computer Science 2 Hunt Alan amp Koplas Geoffrey D 1998 Definitional Vocabulary Acquisition Using Natural Language Processing and a Dynamic Lexicon Department of Computer Science State University of New York at Buffalo 3 Rapaport William J amp Kibby Michael W 2001 Contextual Vocabulary Acquisition Development of a Computational Theory and Educational Curriculum Department of Computer Science and Engineering and Department of Learning and Instruction State University of New York at Buffalo 29
6. demos In order to differentiate the demo files from the background files I label demo files as lt filename gt demo and background files as lt filename gt base where lt filename gt equals any filename All the demo files can be found in my CVA directory at projects rapaport CVA mkb3 CV A demos In addition the background files can be found at projects rapaport CV A mkb3 CVA kbs The completed demos were re created by comparing the abbreviated sample runs that are present in Ehrlich 1995 and by carefully looking at the full version of her sample runs in Appendix 2 of Ehrlich 1995 This appendix can only be found in a computer file that she refers to in her dissertation this appendix file is called appendix two The 23 computer file can be found in my CVA directory at projects rapaport C VA mkb3 CV A docs appendix two Please be aware that appendix two at places seems to have been corrupted over the years It is suggested that anyone trying to re create additional dissertation demos be really careful when looking through this file I would use the abbreviated sample runs that are present in Ehrlich 1995 as a guide to what the expected results of each demo are and appendix two only as a supplement Currently I am working on re creating the hackney demo of Ehrlich 1995 but due to appendix two being corrupted I have not been able to get the expected results that are present in the abbreviated sample runs in Ehrlich 1995
7. entually be taught to students The above is the streamlined version that was developed In addition Scott T Napieralski developed a more detailed English version of the noun algorithm Since this algorithm is around 4 pages it is included in my CVA directory at projects rapaport CVA mkb3 CVA docs Algorithm v1 0 pdf Additionally the streamlined version above can be found in my CVA directory at projects rapaport CV A mkb3 CV A docs Streamlined v1 0 pdf 7 FUTURE WORK One thing that probably should be done is to extend the noun algorithm to recognize the part whole case frame This case frame could be used to report more information in the unified dictionary frame in the slot for structure Also it must be decided what should appear when reporting the definition of a noun It must be decided in what circumstances to report all slots in the definition as opposed to just specific slots Once a particular solution is arrived at the code must be modified to reflect the decided changes Prof Rapaport and I have talked about adding a flag that when set to true reports all slots and when set to false reports only specific slots such specific slots still need to be decided on Another thing that can be done to the code is to add changes proposed by students in Prof Rapaport s Spring 2002 Advanced Knowledge Representation class although 27 various research must be performed to justify these changes made to the algorithm Nex
8. file showing you how to load cselisp el custom set variables user mail address abc zooloo cs buf EDU t query user mail address nil custom set faces load cselisp el Note that this sample emacs file can be found at http www cse buffalo edu mkb3 sample emacs html Also it can be found in my CVA directory at projects rapaport CV A mkb3 CV A instructs sample emacs html Additionally the Instructions for Running CVA Algorithm and SNePS on UB CSE Machines can be found at http www cse buffalo edu mkb3 sneps cva noun inst html Also it can be found in my CVA directory at projects rapaport C V A mkb3 CV A instructs sneps cva noun inst html 2 1 Instructions for Creating a Runnable Demo using The Verb or Noun Algorithm 33 Reset the network resetnet t Don t trace infer setq snip infertrace nil Load all valid relations intext projects rapaport CV A mkb3 CVA rels rels Compose necessary paths so that the Noun Verb Algorithm 335 functions correctly intext projects rapaport CV A mkb3 CVA paths paths 33 Load Karen Ehrlich s Noun Verb Algorithm load projects rapaport C VA mkb3 CV A src fast code Load Knowledge Base intext projects rapaport CV A mkb3 C VA kbs cat base The above is a sample header block that must be present at the beginning of a demo file in order to successfully run a demo using the verb or noun algor
9. he SNePSUL commands it doesn t echo them Note that your path to Ehrlich specific relations may differ from above depending on the location of the rels file The rels file can be found in my CVA directory at projects rapaport CV A mkb3 CVA rels rels This statement is mandatory for the verb and noun algorithms to properly function In addition to loading the rels it is also mandatory to load certain paths that will be used for path based inference within the verb and noun algorithms Similarly to loading the relations we use the intext command to load the paths These paths are important since loading them will ensure that the verb and noun algorithms are correctly functioning Again note that your path to Ehrlich specific paths may differ from above depending on the location of the paths file The paths file can be found in my CVA directory at projects rapaport CV A mkb3 C VA paths paths Next it is absolutely mandatory to load the verb or noun algorithm This is done with the Lisp load function Since the demo is running inside SNePS we can initiate a Lisp call by preceding the load function with a carat It is also recommended that the new fast KE verb or noun algorithm be used fast code as opposed to the slower original verb or noun algorithm code This new fast algorithm will be described in Section 4 below Finally if a knowledge base is needed for the demo you can use the intext command t
10. her focus was on having a reader naturally acquire unknown vocabulary That is to acquire vocabulary without resorting to asking someone or looking it up in a dictionary This algorithm builds a dictionary style definition consisting of specific kinds of information such as actions ownership function and structure I have been slowly doing tasks throughout the current academic year related to improving and enhancing Karen Ehrlich s noun algorithm Ehrlich 1995 By improving I mean more specifically making the noun algorithm more efficient The code contained a great deal of redundancy that has been eliminated In addition I have eliminated a particular error in the algorithm that had caused it to report back incorrect information During the past year I have completed many miscellaneous tasks that were meant to facilitate the coding of several passages by a few students I provided these students with several tools to get started with 1 detailed instructions for running SNePS on our local CSE machines i e all graduate machines 2 a dictionary of case frames for the noun algorithm those that the algorithm actually looks for 3 detailed instructions for actually creating a demo to be run with Ehrlich s verb and noun algorithm 4 a illustrated version of the noun algorithm and 5 recreation of several of Ehrlich s original demos those in Ehrlich 1995 In addition to the above during the Fall 01 semester Professor William
11. included in your CVA passages to run correctly Notice that the demos load the noun algorithm thus if we run a demo it is not necessary to manually load the noun algorithm as above It is also important to note that prefixing anything by a at the SNePS prompt or within a demo file will allow you to call something in Lisp from SNePS It is important to note several things regarding the above instructions First I have chosen to advise students to run an old version of Allegro Common Lisp run old acl or old acl Currently the old version of Allegro Common Lisp ACL on our Computer Science and Engineering local subnet is 5 0 1 Enterprise Edition for SPARC The reason for running an older version of ACL at the current time is a problem with SNePS interacting with the newer version of ACL 6 1 More specifically the problem is that the Lisp image currently shipped with ACL is no longer case insensitive upper A fix is in the works by Fran Johnson current SNWIZ which is in beta testing right now Once this passes beta the new version called by either run acl or acl should be used and these instructions should be updated to reflect this change In addition it is important to note that in order to run both ACL and SNePS in XEmacs or Emacs you must load cselisp el in your emacs file This file can be found in your home directory by typing ls al at the shell prompt The instructions include the following sample emacs
12. ion I ve tried to attach the actual path that each find looks for in the noun algorithm in pictorial format as a hyperlink This is quite huge so I ve placed it in my CVA directory at projects rapaport C V A mkb3 CV A docs code fast pictoral html I should also give credit where credit is due to Scott T Napieralski for allowing me to use his previously created pictures of these paths 22 Although I have succeeded in making the noun algorithm more efficient my solution has continued on some poor Lisp style by Ehrlich pointed out by Professor Shapiro The Lisp setq s that I have used to eliminate most if not all of the redundancy could be a major source of problems in the future This is because setq s define global variables and since these variables are never used anywhere else in the program they should really be defined as local variables using the Lisp let function If someone else in the future works on this code and is unaware of these global variables they could wreak havoc on the algorithm I consider this a very big problem that should be fixed as soon as possible 5 EHRLICH S DEMOS Currently I have been successful in re creating several of Ehrlich s original dissertation demos from Ehrlich 1995 The completed demos include cat stender brachet and tomato In order to run these demos we need some sort of background information so I also re created the relevant background information files for each of these
13. ithm Note that the three semicolons denote a comment and not a SNePSUL or Lisp command The header block contains six statements where some are mandatory and some are optional The first thing that the user should do is reset the network resetnet t This removes all previously built nodes and clears all previous user entered relations defaults back to pre defined SNePS relations only This statement is optional but is highly recommended This will ensure that the user is starting with a clean slate and facilitates discovering possible errors more easily Other nodes that have been previously built during other demos can influence the validity of the dictionary like definition produced by the verb and noun algorithms Next the SNIP variable infertrace is set to nil thus disabling infertrace This statement is optional However I usually choose to set this variable to nil it is set to true by default since it makes a longer demo more easily readable For a further explanation of this SNIP variable refer to the SNePS 2 5 User Manual at http www cse buffalo edu jsantore snepsman The next thing that must be done is to load all relations that are specific to the verb or noun algorithm This is accomplished by using the intext command above The intext command simply reads a sequence of SNePSUL commands from a specified file in our case the paths file and executes each of them As it reads each of t
14. llection rel points to ARE and object2 points to the class in question Reference Sung Hye Cho Representations of collections in a porpositional semantic network In Working Notes of the AAAI 1992 Spring Symposium on Propositional Knowledge Representation AAAI March 1992 19 3 12 subclass superclass subclass superclass Syntax assert subclass i superclass j Semantics m is the proposition that the class of i s is a subclass of the class of j s Sample Context Hounds are dogs describe assert subclass build lex hound superclass build lex dog 3 13 synonym synonym synonym synonym Syntax build synonym i synonym j Semantics m is the proposition that 1 and j are synonyms Sample Context Small and little are synonyms describe assert synonym build lex small synonym build lex little 20 4 A MORE EFFICENT NOUN ALGORITHM I have made the original noun algorithm more efficient It seems that Ehrlich in some of her Lisp functions had some redundancy in her code For instance say we have a function called foo In foo she would have a conditional that tested whether or not a path could be found If the conditional proved to be true she would then again search for the same path and add it to some relevant list To alleviate the redundancy which had an influence on slowing down the algorithm reporting back the definition given some noun as inp
15. nt i act j onto k Semantics m is the proposition that agent 1 performs act j onto k Sample Context King Arthur leaped on a table describe assert object KA proper name build lex King Arthur describe assert member TB class build lex table describe assert agent KA act build lex leap onto TB 11 3 4 antonym antonym antonym antonym Syntax build antonym i antonym j Semantics m is the proposition that i and j are antonyms Sample Context Hot and cold are antonyms describe assert antonym build lex hot antonym build lex cold 12 3 5 member class member class Syntax assert member i class j Semantics m is the proposition that i is a member of class j Sample Context Evelyn is a person describe assert object Eve proper name build lex Evelyn describe assert member Eve class build lex person Required Usage If the class in question is basic level e g table person then you must use the member class case frame and not the object1 rel object2 case frame 13 3 6 mode object mode object Syntax build mode i object j Semantics m is the proposition that the modal concept i is applied to the object j Sample Context Presumably Joe is smart describe assert object Joe proper name build lex Joe describe assert mode build lex presumably obje
16. o load it This is optional and depends on whether you require a knowledge base with your demo Note that your path to the knowledge base may differ from above depending on the location of your knowledge base file All knowledge base files are appended with a base indicating that it is a knowledge base file I suggest that if your demo is called cat demo and you require a knowledge base you call it cat base The knowledge base files for our sample demos can be found in my CVA directory at projects rapaport CV A mkb3 CV A kbs lt filename gt where lt filename gt is the name of the knowledge base file After the header block you can begin representing your passage containing some unknown word As an example look at the cat demo in my CVA directory at projects rapaport CV A mkb3 CVA demos cat demo You will find many other demos in here as well These complete instructions can be found at http www cse buffalo edu mkb3 create demos html In addition they can be found in my CVA directory at projects rapaport CVA mkb3 CV A instructs create demos html 3 A Dictionary of CVA SNePS Case Frames I ve designed a case frame dictionary that includes all the case frames that the noun algorithm searches for when asked to define a noun Since as of now these are the only case frames the algorithm looks for diligence should be used when coding passages where the noun is the unknown word to stay within these case f
17. on must be done to see if removing demo dependencies makes any difference 25 6 STREAMLINED AND MORE DETAILED ENGLISH VERSION OF KE NOUN ALGORITHM During the fall semester of 2001 my first task along with Professor William J Rapaport defn noun N generate a definition for a noun N Report the following information about Ns if known 1 class membership else names of individuals a in general report the 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 if there are no known class memberships then report names of individuals who are Ns properties else possible properties of Ns a if there are no known class memberships i e if only individual Ns are known then report their properties as possible properties of Ns structural else possible structural information functions else possible functions of Ns acts else possible acts that Ns perform agents that perform acts on Ns and the acts they perform ownership information synonyms and Scott T Napieralski was to develop a streamlined English version of the noun 26 algorithm The hope was that while creating such a version we would learn how the algorithm worked and be able to supply the education members of our research group with a beginning point in developing a curriculum to ev
18. rame restraints The knowledge engineer is restricted to these thirteen case frames 1 agent act 2 agent act object 3 agent act onto 4 antonym antonym 5 member class 6 mode object 7 object proper name 8 object property 9 object rel possessor 10 object1 rel object2 11 objects1 rel object2 12 subclass superclass and 13 synonym synonym A complete listing of the syntax semantics for these case frames can be found in the KE Case Frame Dictionary at http www cse buffalo edu mkb3 case frame dictionary case frame index html In addition it can also be found in my CVA directory at projects rapaport CV A mkb3 CV A dictionary case frame index html 3 1 agent act agent act Syntax build agent i act j Semantics m is the proposition that agent i performs act j Sample Context Joe sleeps describe assert object Joe proper name build lex Joe describe assert agent Joe act build lex sleep 3 2 agent act object agent act object Syntax build agent i act j object k Semantics m is the proposition that agent i performs act j with repect to k Sample Context Joe hits the ball describe assert object Joe proper name build lex Joe describe assert member Ball class build lex ball describe assert agent Joe act build lex hit object Ball 10 3 3 agent act onto agent act onto Syntax build age
19. t the Lisp setq s described above in Section 4 must be changed to Lisp let s to turn global variables to local variables The reasons are described above although the more I ve discussed such changes with Prof Rapaport the more we think that it may be more practical to completely rewrite the noun algorithm There seem to be a variety of errors we are experiencing and a lack of style and efficiency in the current algorithm By continuing to work on this with the hope of further enhancing it we may be compounding the errors already present So it is my opinion that the algorithm should be rewritten if so chosen by Prof Rapaport and the above changes in this Section could be built in from scratch Another important thing that must be done is resurrect SNePSwD from the dead It appears from my conversation with Prof Rapaport and reading Hunt and Koplas 1998 that a particular problem with trying to get SNePSwD back and running was the new updated version of Allegro Common Lisp That is all the information I have found to give on trying to get SNePSwD up and running again This is important since in order to test some of the verb demos in Ehrlich 1995 we must get the belief revision up and running once again Although if we chose to rewrite the algorithm it would be more practical to try to incorporate a new belief revision system currently being worked on by Fran Johnson in our department under the guidance of Prof Stuart Shapiro instea
20. ut the choice was made to save the path to a Lisp variable while inside the conditional If this variable turns out to contain something other then nil we can just add it to the relevant list without searching for it again Take a look at the following snippet of code before such changes were made and then after definite rule or entail basic ctgy transitive basic object cond AND 3 find compose lex act cq ant class lex noun compose lex act agent member class lex noun 3 find compose lex class member object cq ant class lex noun list 3 find compose lex act cq ant class lex noun compose lex act agent member class lex noun 3 find compose lex class member object cq ant class lex noun Before 21 definite rule or entail basic ctgy transitive basic object cond AND setq partl 3 find compose lex act eq ant class lex noun compose lex act agent member class lex noun setq part2 3 find compose lex class member object cq ant class lex noun list parti part2 After The new fast code can be found in my CVA directory at projects rapaport CV A mkb3 CVA src fast code I ve also left the old code in the same directory at projects rapaport CV A mkb3 CVA sre code In order to aid future workers who need to work on this new fast algorithm I have created an illustrated vers
21. velyn describe assert object Pye rel build lex cat possessor Eve 17 3 10 objectl1 rel object2 object1 rel object2 Syntax assert object1 i rel j object2 k Semantics m is the proposition that j i k Sample Context Jack is next to Fred describe assert object Jack proper name build lex Jack describe assert object Fred proper name build lex Fred describe assert object1 Jack rel build lex next to object2 Fred Special Usage class inclusion If the class in question is either a subordinate e g coffee table or a superordinate e g furniture then the object1 rel object2 case frame must be used where object points to some individual rel points to ISA and object2 points to the class in question 18 3 11 objects1 rel object2 objects1 rel object2 Syntax assert objects i rel j object2 k Semantics m is the proposition that j i k Sample Context Knights varled their shields describe assert agent KS act build lex varl object SH time KVS describe assert objects KS rel build lex are object2 build lex knight describe assert members SH class build lex shield Special Usage If the class in question is either a subordinate e g coffee table or a superordinate e g furniture then the objects1 rel object2 case frame must be used where objects points to some definite plural entity co
Download Pdf Manuals
Related Search