Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive...
Transcript of Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive...
![Page 1: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/1.jpg)
Inductive Detection of Language Features via Clustering Minimal Pairs
Toward Feature-Rich Grammars inMachine Translation
Jonathan ClarkRobert Frederking
Lori Levin
Language Technologies InstituteCarnegie Mellon University
Pittsburgh, PA
ACL SSST-2 - June 20, 2008
![Page 2: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/2.jpg)
• Overview of Feature Detection
• Example Application: Feature-Rich Grammars
• The Process of Feature Detection
• Results
• Conclusions
Outline
![Page 3: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/3.jpg)
Feature Detection((num sg)…)The dog eatsEl perro come
((num dl)…)The dogs eatLos perros comen
((num pl)…)The dogs eatLos perros comen
![Page 4: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/4.jpg)
Feature Detection((num sg)…)The dog eatsEl perro come
((num dl)…)The dogs eatLos perros comen
((num pl)…)The dogs eatLos perros comen
context: There are two dogs
![Page 5: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/5.jpg)
Feature Detection((num sg)…)The dog eatsEl perro come
((num dl)…)The dogs eatLos perros comen
((num pl)…)The dogs eatLos perros comen
![Page 6: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/6.jpg)
Feature Detection((num sg)…)The dog eatsEl perro come
((num dl)…)The dogs eatLos perros comen
((num pl)…)The dogs eatLos perros comen
Does this language distinguish singular, dual, or plural agents? If so, how?
![Page 7: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/7.jpg)
Feature Detection
Feature Candidate Lexical Items
(num sg) come, el, perro
(num dl, pl) comen, los, perros
FeatureDetection
((num sg)…)The dog eatsEl perro come
((num dl)…)The dogs eatLos perros comen
((num pl)…)The dogs eatLos perros comen
Does this language distinguish singular, dual, or plural agents? If so, how?
![Page 8: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/8.jpg)
✓Overview of Feature Detection
• Example Application: Feature-Rich Grammars
• The Process of Feature Detection
• Results
• Conclusions
Outline
![Page 9: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/9.jpg)
Problem: Long-Distance Interactions
“A student named Irshad who was throwing flour in the water for the fish . . . ”
←12 words→
English → Urdu
![Page 10: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/10.jpg)
...
Example Application: Feature-Rich Grammars
←12 words→
![Page 11: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/11.jpg)
...
Example Application: Feature-Rich Grammars
NPA VP
←12 words→
![Page 12: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/12.jpg)
...
Example Application: Feature-Rich Grammars
NPA VP
(num sg) (num sg)
←12 words→
![Page 13: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/13.jpg)
...
Example Application: Feature-Rich Grammars
NPA VP
S
(NPA num) = (VP num)
(num sg) (num sg)
←12 words→
![Page 14: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/14.jpg)
...
Example Application: Feature-Rich Grammars
NPA VP
S
(NPA num) = (VP num)
(num sg) (num sg)
S
(num pl)
(NPA num) = (VP num)
←12 words→
![Page 15: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/15.jpg)
✓Overview of Feature Detection
✓Example Application: Feature-Rich Grammars
• The Process of Feature Detection
• Results
• Conclusions
Outline
![Page 16: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/16.jpg)
Inductive Feature Detection
Minimal Pair Clustering
Feature Value Clustering
Morpheme-Feature Pairing
Elicitation Corpus
Morpheme-Feature Pairs
![Page 17: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/17.jpg)
Elicitation Corpus Entrycontext: Maria bakes cookies regularly or habitually. srcsent: Maria bakes cookies .
![Page 18: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/18.jpg)
Elicitation Corpus Entrycontext: Maria bakes cookies regularly or habitually. srcsent: Maria bakes cookies .
![Page 19: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/19.jpg)
Elicitation Corpus Entrycontext: Maria bakes cookies regularly or habitually.srcsent: Maria bakes cookies .tgtsent: Maria hornea galletas . aligned: ((1,1),(2,2),(3,3),(4,4))
![Page 20: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/20.jpg)
Elicitation Corpus Entrycontext: Maria bakes cookies regularly or habitually.srcsent: Maria bakes cookies .tgtsent: Maria hornea galletas . aligned: ((1,1),(2,2),(3,3),(4,4))fstruct: [f1]( [f2](actor ((gender f)(anim human)(num sg))) [f3](undergoer ((person 3) (num dl))) (tense pres))
![Page 21: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/21.jpg)
Elicitation Corpus Entrycontext: Maria bakes cookies regularly or habitually.srcsent: Maria bakes cookies .tgtsent: Maria hornea galletas . aligned: ((1,1),(2,2),(3,3),(4,4))fstruct: [f1]( [f2](actor ((gender f)(anim human)(num sg))) [f3](undergoer ((person 3) (num dl))) (tense pres))
Distributed in this year’s NIST Urdu MT Evaluation via the LDC’s LCTL Language Pack
![Page 22: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/22.jpg)
Inductive Feature Detection
Minimal Pair Clustering
Feature Value Clustering
Morpheme-Feature Pairing
Elicitation Corpus
Morpheme-Feature Pairs
![Page 23: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/23.jpg)
Minimal Pair Clustering
pl
The dogs see the catLos perros ven el gato((actor (num pl)(gender n)) (undergoer (num sg)(gender n)) (tense pres))
The dog sees the catEl perro ve el gato((actor (num sg)(gender n)) (undergoer (num sg)(gender n)) (tense pres))
sg
Wildcard Feature Structure:((actor (num *)(gender n)) (undergoer (num sg)(gender n)) (tense pres))
![Page 24: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/24.jpg)
Minimal Pair Clustering
pl
The dogs see the catLos perros ven el gato((actor (num pl)(gender n)) (undergoer (num sg)(gender n)) (tense pres))
The dog sees the catEl perro ve el gato((actor (num sg)(gender n)) (undergoer (num sg)(gender n)) (tense pres))
sg
Wildcard Feature Structure:((actor (num *)(gender n)) (undergoer (num sg)(gender n)) (tense pres))
![Page 25: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/25.jpg)
Inductive Feature Detection
Minimal Pair Clustering
Feature Value Clustering
Morpheme-Feature Pairing
Elicitation Corpus
Morpheme-Feature Pairs
![Page 26: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/26.jpg)
Feature Value Clusterings1: The dog sees the cat El perro ve el gato ((actor (num sg)…))
sg
![Page 27: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/27.jpg)
Feature Value Clusterings1: The dog sees the cat El perro ve el gato ((actor (num sg)…))
sg
s2:The dogs see the catLos perros ven el gato((actor (num dl)…))
{ (s1, s2, NEQ) }
dl
![Page 28: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/28.jpg)
Feature Value Clustering
pl s3: The dogs see the catLos perros ven el gato((actor (num pl)…))
{ (s1, s3, NEQ) }
{ (s1, s3, EQ) }
s1: The dog sees the cat El perro ve el gato ((actor (num sg)…))
sg
s2:The dogs see the catLos perros ven el gato((actor (num dl)…))
{ (s1, s2, NEQ) }
dl
![Page 29: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/29.jpg)
Feature Value Clustering
pl s3: The dogs see the catLos perros ven el gato((actor (num pl)…))
{ (s1, s3, NEQ) }
{ (s1, s3, EQ) }
s1: The dog sees the cat El perro ve el gato ((actor (num sg)…))
sg
s2:The dogs see the catLos perros ven el gato((actor (num dl)…))
{ (s1, s2, NEQ) }
dl
![Page 30: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/30.jpg)
Feature Value Clustering
dl-pl
s3: The dogs see the cat Los perros ven el gato ((actor (num dl)…))
{ (s1, s3, EQ) }
s1: The dog sees the cat El perro ve el gato ((actor (num sg)…))
s2: The dogs see the cat Los perros ven el gato ((actor (num dl)…))
{ (s1, s2, NEQ), (s1, s3, NEQ) }
sg
![Page 31: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/31.jpg)
Inductive Feature Detection
Minimal Pair Clustering
Feature Value Clustering
Morpheme-Feature Pairing
Elicitation Corpus
Morpheme-Feature Pairs
![Page 32: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/32.jpg)
Morpheme-Feature Pairing
dl-pl
s3: The dogs see the cat Los perros ven el gato ((actor (num dl)…))
{ (s1, s3, EQ) }
s1: The dog sees the cat El perro ve el gato ((actor (num sg)…))
s2: The dogs see the cat Los perros ven el gato ((actor (num dl)…))
{ (s1, s2, NEQ), (s1, s3, NEQ) }
sg
![Page 33: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/33.jpg)
Morpheme-Feature Pairing
dl-pl
s3: The dogs see the cat Los perros ven el gato ((actor (num dl)…))
{ (s1, s3, EQ) }
s1: The dog sees the cat El perro ve el gato ((actor (num sg)…))
s2: The dogs see the cat Los perros ven el gato ((actor (num dl)…))
{ (s1, s2, NEQ), (s1, s3, NEQ) }
sg
{ el, perro, ve } { los, perros, ven }
![Page 34: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/34.jpg)
✓Overview of Feature Detection
✓Example Application: Feature-Rich Grammars
✓The Process of Feature Detection
• Results
• Conclusions
Outline
![Page 35: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/35.jpg)
Evaluation
Minimal Pair Clustering
Feature Value Clustering
Morpheme-Feature Pairing
Elicitation Corpus
Morpheme-Feature Pairs
![Page 36: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/36.jpg)
Experiment• Analyzed LCTL Urdu-English Elicitation Corpus
(~3000 sentences)
• Evaluated by Urdu native speaker knowledgeable in linguistics
Judgement Morpheme-Feature Pairings
ExampleOutput
Correct 68 “hai” → (tense pres)
Ambiguous 29 “raha” → (tense pres)
Incorrect 109 “nahin” → (tense pres)
TOTAL 206
![Page 37: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/37.jpg)
Experiment
• Found 68 Correct Morpheme-Feature Pairs
• = 53 Word Types
• In an Urdu corpus of 17M tokens and 200k types, the 53 types selected by feature detection cover:
• 0.02% of Types (Vocabulary)
• 38.6% of Tokens
61%
39%
AnnotatedRemaining
Tokens inBlind Test Set
![Page 38: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/38.jpg)
✓Overview of Feature Detection
✓Example Application: Feature-Rich Grammars
✓The Process of Feature Detection
✓Results
• Conclusions
Outline
![Page 39: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/39.jpg)
Minimal Pair Clustering
Feature Value Clustering
Elicitation Corpus Future Work
Morpheme-Feature Pairing
Morpheme-Feature Pairs
Morphological Segmentation
![Page 40: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/40.jpg)
Minimal Pair Clustering
Feature Value Clustering
Elicitation Corpus Future Work
Morpheme-Feature Pairing
Morpheme-Feature Pairs
Morphological Segmentation
Presented MethodFeature Candidate Lexical Items
(num sg) come, el, perro
(num dl, pl) comen, los, perros
Future WorkFeature Candidate Morphemes
(num sg) VB+Ø, el, N+o
(num dl, pl) VB+n, los, N+os
![Page 41: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/41.jpg)
Minimal Pair Clustering
Feature Value Clustering
Elicitation Corpus Future Work
Morpheme-Feature Pairing
Morpheme-Feature Pairs
Morphological Segmentation
![Page 42: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/42.jpg)
Minimal Pair Clustering
Feature Value Clustering
Elicitation Corpus Future Work
Morpheme-Feature Pairing
Morpheme-Feature Pairs
Filtering
Morphological Segmentation
![Page 43: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/43.jpg)
Minimal Pair Clustering
Feature Value Clustering
Elicitation Corpus Future Work
Morpheme-Feature Pairing
Morpheme-Feature Pairs
Filtering
Feature Constraint Learning
Morphological Segmentation
![Page 44: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/44.jpg)
Minimal Pair Clustering
Feature Value Clustering
Elicitation Corpus Future Work
Morpheme-Feature Pairing
Morpheme-Feature Pairs
Filtering
Feature Constraint Learning
Structural MT Integration
Morphological Segmentation
![Page 45: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/45.jpg)
Other Applications
• Factored MT
• Data Selection via active learning for synchronous grammar induction
• Aid for linguistics field work
![Page 46: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/46.jpg)
Conclusion
• We now have
• Feature-annotations for lexical items that convey grammatical meanings
• Significant coverage
• Structural MT systems stand to benefit by incorporating this morphosyntactic information
![Page 47: Inductive Detection of Language Features via Clustering ...jhclark/pubs/clark_ssst2.pdfInductive Detection of Language Features via Clustering Minimal Pairs Toward Feature-Rich Grammars](https://reader033.fdocuments.in/reader033/viewer/2022042201/5ea1788f081aca646a26baf3/html5/thumbnails/47.jpg)
Inductive Detection of Language Features via Clustering Minimal Pairs:
Toward Feature-Rich Grammars in Machine Translation
Jonathan ClarkRobert Frederking
Lori Levin
Language Technologies InstituteCarnegie Mellon University
Pittsburgh, PA
Questions?