Unlocking and Sharing LTCL Linguistic Knowledge

Download Report

Transcript Unlocking and Sharing LTCL Linguistic Knowledge

With 6,500 languages in the world,
we must explore
new ways to learn, document, and share
our linguistic knowledge.
John J. Kovarik
NSA/CSS Senior Language Technology Authority
Unlocking and Sharing LTCL
Linguistic Knowledge
Keywords: CFG parsing, language generation,
computational linguistics
CALICO ’05
University of Michigan
Ann Arbor, MI May 17-20, 2005
The Challenges of Learning
and Sharing Knowledge of an
LCTL in the 21st Century
John J. Kovarik
National Security Agency
Presentation Overview
 General LCTL Challenges
 Challenges of Learning Mongolian
 Recipe for New Approach
 Khalka Mongolian Parts of Speech
 Mongolian Morphological Affixes
 Method of Lexical Knowledge Representation
 Analyze, Parse, Build Grammar Model, Test
 Iterate Repeatedly
LCTL Learning Challenges
 Fewer Learned Resources to Learn from
 Less Recognition Nationally
 Less Opportunities to Document What’s Learned
 Very Few Students to Learn from You
 Almost All Learning Done Manually
 Few Reliable 21st Century Applications
– Microsoft IME
– Font
Mongolian Learning Challenges
 Input Method Emulator (IME)
– MicroSoft IME
• Keyboard arranged for native Mongols
• American Mongolists prefer phonetic keyboard
– “a” key on Mongolian keyboard mapped to ASCII “a” etc.
 Fonts commonly used on Internet
– Russian Cyrillic fonts are commonly used
• “|” and “0” commonly substituted for “ү” and “ө”
• “у” and “о” often freely extended to “ү” and “ө”
Recipe for a New Approach
 Take a student with a computational linguistics
background
 Infuse with curiosity and energy
 Stir in access to the Internet
 Add Mongolian syntax and morphology
 Create morphological analyzer, context free
parser, and grammatical generator for Mongolian
 Resulting lexicons, software, and grammar models
can be used by other linguistically adept students
Khalkha Mongolian
Parts of Speech
 Declinable Nouns
 Declinable Adjectives
 Inflected Verbs
 Unchanging Adverbs
 Declinable Converbs
 Unchanging Postpositions
 Unchanging Conjunctions
 Unchanging Particles
Mongol Morphological Affixes
 27 verbal suffixes denoting tense and mood
 2 verb infixes denoting verb manner
– Consultative
– Passive
 6 verb paradigms or verb types
 3 irregular common verbs
 6 cases in singular and plural number
 Both nouns and adjectives are declined
Lexical Knowledge Representations
 Unchanging adverbs, conjunctions,
particles, etc. and irregular verb forms
(unchanging.txt file)
 Lemmas of declinable nouns and adjectives
(declinables.txt file)
 Inflected verbs and nominalized verbs
(regvb.txt file)
 Affix files (casendings.txt, reflex.txt,
infixes.txt, vbforms.txt)
Some Examples
 declinables.txt file
– N нэр Q хэн
 regverb.txt file
– V ир
V өс
 Affix files
– casendings.txt g ний d д
a ыг b оос
– reflex.txt
аа
ээ
оо
– infixes.txt
C лц R лд P гд
– vbforms.txt)
ipf нө i1p в i3p чээ Ypf охгүй
 unchanging.txt file
– Pg->талаар
Pc->холбогдуулан
Merge Morphology Knowledge
with the Power of the Computer
Wrote yalgah.pl to become tireless lexical pedagogue
 Searches for identifiable affixes by comparison
with lexical knowledge affix files
 Matches resulting lemma against lexical
knowledge declinables, verbs, and unchanging
words, then outputs word/part of speech tag to
standard output file plus expository lexicon
 Depending whether lemma can or cannot be
matched, outputs:
• Lemma to Out Of Vocabulary (oov) file noting affixes found
• Word/part of speech tag to standard output file
Additional Outputs
 Expository Morphology File (named morphlex.txt)
IR->verb command imperative 2nd person singular
IREEREY->converb future perfect continuative
IREG-> verb command concessive 3rd person singular/plural
BAGA->adjective
HURAL->noun nominative
IH->adjective
AJILDAA->reflexive noun dative-locative
ORLOO->verb indicative second past

Out Of Vocabulary File (named oov)
[C = : = > 5 = 0 E 0 0 A 0 0 ] (UNKNOWNAHAASAA) WORD 0 LINE 2
FALLS OUTSIDE OF VOCABULARY
possible reflexive ending <0 0 >-<AA>
possible declinable case ending<b>-<0 0 A >-<AAS>
possible verbal part of speech <Ypf >-<0 E >-<AH>
possible participial/converbal stem <C = : = > 5 = >--<UNKNOWN>
Feed Analytic Output to Parser

Developed context-free grammar (CFG) rules for both discourse and newspaper texts
S->Sbj Prd
S->Prd
Sbj->Nn Sbj->NP
NP->Tg Nn
NP->Tg Ng Nn
Prd->J


Wrote parse.pl to validate CFG rules against input text tagged as to part of speech
When each sentence can be fully parsed, outputs a parse tree and an English gloss.
Working on "BAGA HURAL IH AJILDAA ORLOO ."
ENGLISH GLOSS: large hural great work began .
The sentence does parse.
Branch nodes on tree:
S -> (Sbj Prd)
Sbj -> (NP)
NP -> (J Nn)
Prd -> (NPd Vi2p)
NPd -> (J Nd)
POS: J Nn J Nd Vi2p
Feed Output to Generator
 Wrote gramgen.pl to generate sentences
based on lexical knowledge, morphological
knowledge, and syntactic knowledge gained
 Output routinely reviewed for accuracy and
Chomskian explanatory adequacy of the
grammar models created for the parser and
generator engines
Iterative Process
 First take new newspaper article or dialogue and
run morphological analyzer on it until all words
are listed within vocabulary (no output in the oov
[Out Of Vocabulary] file
 Run output through parser, creating new CFG
rules until new text parses
 Run generator for a hundred or more examples to
ensure adequacy of new rules
Morpho-analyzer, Parser, Generator
Software Led This Student to Deeper
Understanding of Mongolian
 A linguistically adept learner can thus write
software to help one learn deeper & faster
 Language tool development is thus
grounded in gaining and applying language
knowledge in a systematic and linguistically
principled manner for oneself and others
Contact Information
John Kovarik
 Email: [email protected]
 Home Page:
http://www.worldnet.att/~kovariks
 Phone: 443-479-7188