Rules and analogy in Russian Rules and analogy in Russian loanword adaptation and loanword adaptation and
novel verb formationnovel verb formation
Vsevolod KapatsinskiIndiana University
Dept. of Linguistics &Cognitive Science Program
Speech Research [email protected]
LSA 2007
Russian stem extensionsRussian stem extensions
-i- event -i- event event+i+ ‘happen’ event+i+ ‘happen’ -a- eat -a- eat it+a+ ‘eat’ it+a+ ‘eat’
Source: Source: The BigThe Big Dictionary of Youth Dictionary of Youth Slang, Slang, 20032003 Borrowed verbsBorrowed verbs New verbs formed from nounsNew verbs formed from nouns
Which stem extensions are Which stem extensions are more productive?more productive?
Relative productivity of Russian stem extensions
0
100
200
300
400
i a ova nu e irova nicha
Extension
Num
ber o
f new
ver
bs
The questionsThe questions How can we predict the choice of the stem How can we predict the choice of the stem
extension?extension?
Is one extension applied by default? Is one extension applied by default? Predicted by the Dual Mechanism Model (Pinker and Predicted by the Dual Mechanism Model (Pinker and
Prince 1988, 1994)Prince 1988, 1994)
Locality effectsLocality effects Analogical vs. schema-based accounts?Analogical vs. schema-based accounts?
Do parts of the root adjacent to the root-suffix boundary Do parts of the root adjacent to the root-suffix boundary influence suffix choice more than more distant parts of influence suffix choice more than more distant parts of the root?the root?
Do parts of the root that are not adjacent to the root-Do parts of the root that are not adjacent to the root-suffix boundary influence the choice of the suffix?suffix boundary influence the choice of the suffix?
Unexpected under the Rule-Based Learner (Albright Unexpected under the Rule-Based Learner (Albright and Hayes 2003)and Hayes 2003)
Phonotactic influences:Phonotactic influences:It’s not all phonotacticIt’s not all phonotactic
0
20
40
60
80
100
Labials Coronals Velars
[Place] of root-final C
% o
f ve
rbs
taki
ng
-a
vs.
-i
a
i
Phonotactics do not Phonotactics do not explain all the variationexplain all the variation
Can analogy to existing words predict Can analogy to existing words predict the stem extension taken by a the stem extension taken by a borrowed verb?borrowed verb?
Analogy:Analogy: The borrowed verb will take the stem The borrowed verb will take the stem
extension of the majority of its neighbors.extension of the majority of its neighbors. Verbs are neighbors if their roots share Verbs are neighbors if their roots share
at least 2/3 of their phonemesat least 2/3 of their phonemes
Analogical predictionsAnalogical predictions
kam
kap
kak
kaz
kar
kaj
kad
kim
xam
kum
kajm
kach
-a
-a
-i
Similarity effectSimilarity effect
N=598
N=1085
0%
20%
40%
60%
80%
100%
most neighbors bear a most neighbors bear i
verb bears a
verb bears i
i
a
Final consonant as a predictorFinal consonant as a predictor
KAM
kajMxaMkuM
groMtoM
weMshtorMskoroM
KiMduM
xroM
8/11
3/11
m i
Not just Place:b i (41/54)p a (36/57)
Analogy vs. Final Analogy vs. Final consonantconsonant
Breakdown by stem Breakdown by stem extensionextensionAnalogy performs better than final consonant
-a is less predictable than -i based on analogy
0%
20%
40%
60%
80%
100%
coronal labial
Place of articulation of the final consonant
Pe
rce
nt
co
rre
ct
analogy -i
final C -i
analogy -a
final C -a
When analogy makes no When analogy makes no predictionprediction
In 8.5% of verbs, analogy makes no In 8.5% of verbs, analogy makes no predictionprediction Numbers of nieghbors taking each Numbers of nieghbors taking each
stem extension are equalstem extension are equal
OROR No neighborsNo neighbors
What determines stem extension choice What determines stem extension choice then?then?
When number of neighbors taking -i is the same as the number of neighbors taking -a
01020
304050
coronal-final labial-final velar-final
Nu
mb
er o
f ve
rbs
form
ed w
ith
a
i
N=98 (5.5%)
•When there are equal numbers of neighbors rooting for –a and -i, coronals are not associated with either stem extension
•What about verbs that have no neighbors?
Number of neighbors=0Number of neighbors=0
0
10
20
30
40
50
60
coronal-final labial-final velar-finalNu
mb
er
of
ve
rbs
fo
rme
d
wit
h a
i
N=59 (3%)
When there are no neighbors, coronals are always followed by -i
Interim SummaryInterim Summary Analogy accounts for 87% of the data Analogy accounts for 87% of the data
excluding velarsexcluding velars
Analogy performs better than specifying the Analogy performs better than specifying the final consonantfinal consonant
Analogy predicts –i better than it predicts –a Analogy predicts –i better than it predicts –a (70% vs. 93%)(70% vs. 93%)
When there are no neighbors, coronals are When there are no neighbors, coronals are always followed by -ialways followed by -i
An issue for the Dual An issue for the Dual Mechanism ModelMechanism Model
Pinker and Prince (1988, 1994):Pinker and Prince (1988, 1994): One suffix should be more productive than the other One suffix should be more productive than the other
suffix with novel lexical items that are not similar to suffix with novel lexical items that are not similar to existing onesexisting ones
-i > –a after coronals -i > –a after coronals -i is the default-i is the default
This suffix is applied by default. Hence, analogy should This suffix is applied by default. Hence, analogy should be less able to predict when this suffix will occur.be less able to predict when this suffix will occur.
Analogy is less able to predict occurrence of –aAnalogy is less able to predict occurrence of –a -a is the default -a is the default
Possible accounts:Possible accounts: AnalogyAnalogy Associations between parts of the root and suffixesAssociations between parts of the root and suffixes
Associations should be stronger when the distance between the Associations should be stronger when the distance between the suffix and the part of the root is smallsuffix and the part of the root is small
Do neighbors that don’t Do neighbors that don’t share the final C matter?share the final C matter?
Albright and Hayes (2003):Albright and Hayes (2003): The only segment strings that can be The only segment strings that can be
associated with a suffix are uninterrupted associated with a suffix are uninterrupted segment strings that include the final segment strings that include the final segmentsegment
Weaker version:Weaker version: Suffixes can be associated with adjacent Suffixes can be associated with adjacent
phonological chunks more strongly than phonological chunks more strongly than with non-adjacent oneswith non-adjacent ones
Testing the hypothesis of lack of Testing the hypothesis of lack of non-local dependenciesnon-local dependencies
KAM
KAp
KAk
KAz
KAr
KAj
KAd
KiM
xAM
KuM
KAjM
KAch
-a
-a
-i
Adjacent dependencies Adjacent dependencies are strongerare stronger
When neighbors sharing final C and neighbors not sharing final C make different predictions
0
20
40
60
80
neighbors that share finalconsonant
neighbors that DO NOT share thefinal consonant
Nu
mb
er o
f ve
rbs
that
go
wit
h
X
Combining predictorsCombining predictors
If we knowIf we know What do most neighbors sharing final C What do most neighbors sharing final C
take?take? What do most words with this final C take?What do most words with this final C take?
Do we need to knowDo we need to know What do most neighbors that do not share What do most neighbors that do not share
final C take?final C take?
Final consonant vs. final-sharing Final consonant vs. final-sharing neighborsneighbors
KAM
KiM
XaM
KuM
KAjM
loMgroMweMgreMEtc.
Previously sharing just the final C was not enough to be considered neighborsPreviously sharing just the final C was not enough to be considered neighbors
Non-local dependencies Non-local dependencies still importantstill important
Logistic Regression:Logistic Regression: Final C: χFinal C: χ22= 31.0= 31.0 Neighbors sharing final C: χNeighbors sharing final C: χ22 = 329.8 = 329.8 Neighbors not sharing final C: χNeighbors not sharing final C: χ22 = =
181.7181.7
Local dependencies are strongerLocal dependencies are stronger
All predictors are significant at p<.0005All predictors are significant at p<.0005
Non-local dependencies do existNon-local dependencies do exist
ConclusionConclusion
Huge similarity effects for both stem Huge similarity effects for both stem extensionsextensions
All productive suffixes sensitive to All productive suffixes sensitive to similaritysimilarity
But, after coronalsBut, after coronals -a is less predictable than –i based on analogy-a is less predictable than –i based on analogy -i is more productive than –a when there are -i is more productive than –a when there are
no analogical models nearbyno analogical models nearby
Defining attributes of a DMM default are Defining attributes of a DMM default are dissociable (cf. Kapatsinski 2005)dissociable (cf. Kapatsinski 2005)
ConclusionConclusion
-a is less predictable than –i based on -a is less predictable than –i based on analogyanalogy
Possible reason: Possible reason: There are more –i verbs than –a verbs in There are more –i verbs than –a verbs in
the lexiconthe lexicon Possible analogical solution:Possible analogical solution:
Thus, a given neighbor is more likely to Thus, a given neighbor is more likely to bear –i than it is to bear –abear –i than it is to bear –a
Thus, occurrence of an –a neighbor is more Thus, occurrence of an –a neighbor is more salient than occurrence of an –i neighborsalient than occurrence of an –i neighbor
ConclusionConclusion
After coronalsAfter coronals -i is more productive than –a when there -i is more productive than –a when there
are no analogical models nearbyare no analogical models nearby -i and –a are equally productive when -i and –a are equally productive when
there are as many neighbors bearing –i there are as many neighbors bearing –i as neighbors bearing -aas neighbors bearing -a
Interpretation:Interpretation: Use analogy whenever possible; Use analogy whenever possible; if both alternatives have equal support, then if both alternatives have equal support, then
they are equally acceptable; they are equally acceptable; if no analogical models, use phonotacticsif no analogical models, use phonotactics
ConclusionConclusion Analogy or schemas?Analogy or schemas?
Activate similar words?Activate similar words? Activate sublexical chunks associated with suffixes?Activate sublexical chunks associated with suffixes?
Locality effects support the schematic account Locality effects support the schematic account (cf. Albright and Hayes 2003)(cf. Albright and Hayes 2003):: Dependencies between adjacent segments are Dependencies between adjacent segments are
easier to learn than dependencies between non-easier to learn than dependencies between non-adjacent ones (e.g., Hudson Kam and Newport adjacent ones (e.g., Hudson Kam and Newport 2005)2005)
While adjacent dependencies are stronger, non-While adjacent dependencies are stronger, non-adjacent dependencies seem to also play a role in adjacent dependencies seem to also play a role in suffix choice (contra Albright and Hayes 2003).suffix choice (contra Albright and Hayes 2003).
AcknowledgementsAcknowledgements
N.I.H. for financial support through N.I.H. for financial support through a training grant to David Pisoni and a training grant to David Pisoni and the Speech Research Labthe Speech Research Lab
Tessa Bent, Adam Buchwald, Joan Tessa Bent, Adam Buchwald, Joan Bybee, and Susannah Levi for Bybee, and Susannah Levi for helpful discussionhelpful discussion
ReferencesReferences Albright, A., and B. Hayes. 2003. Rules vs. analogy in English past tenses: A Albright, A., and B. Hayes. 2003. Rules vs. analogy in English past tenses: A
computational/ experimental study. computational/ experimental study. CognitionCognition 90, 119-61. 90, 119-61. Bybee, J. L. 1985. Bybee, J. L. 1985. Morphology: A study of the relation between meaning and Morphology: A study of the relation between meaning and
formform. Benjamins.. Benjamins. Bybee, J. L. 1995. Regular morphology and the lexicon. Bybee, J. L. 1995. Regular morphology and the lexicon. Language and Language and
Cognitive ProcessesCognitive Processes,, 10. 425-455.10. 425-455. Kapatsinski, V. M. 2005. Characteristics of a rule-based default are Kapatsinski, V. M. 2005. Characteristics of a rule-based default are
dissociable: Evidence against the Dual Mechanism Model. In S. Franks, F. dissociable: Evidence against the Dual Mechanism Model. In S. Franks, F. Y. Gladney, and M. Tasseva-Kurtchieva, eds. Y. Gladney, and M. Tasseva-Kurtchieva, eds. Formal Approaches to Slavic Formal Approaches to Slavic Linguistics 13: The South Carolina MeetingLinguistics 13: The South Carolina Meeting, 136-46. Michigan Slavic , 136-46. Michigan Slavic Publications.Publications.
Pinker, S., and A. Prince. 1988. On language and connectionism: Analysis of Pinker, S., and A. Prince. 1988. On language and connectionism: Analysis of a parallel distributed processing model of language acquisition. a parallel distributed processing model of language acquisition. CognitionCognition, , 28, 73-193.28, 73-193.
Pinker, S., and A. Prince. 1994. Regular and irregular morphology and the Pinker, S., and A. Prince. 1994. Regular and irregular morphology and the psychological status of rules of grammar. In S. D. Lima, R. L. Corrigan, and psychological status of rules of grammar. In S. D. Lima, R. L. Corrigan, and G. K. Iverson, eds. G. K. Iverson, eds. The reality of linguistic rulesThe reality of linguistic rules, 321-51. Benjamins., 321-51. Benjamins.
Breakdown by place of Breakdown by place of articulation of final Carticulation of final C
0%10%20%30%40%50%60%70%80%90%
100%
coronal labial velar
Final consonant is
Per
cen
t o
f ca
ses
in w
hic
h
anal
og
y is
co
rrec
t vs
. w
ron
g
correct
wrong
Extracting the Extracting the dependenciesdependencies
For a dependency between a part of the For a dependency between a part of the root and a suffix to be formed, many roots root and a suffix to be formed, many roots must share the same sublexical chunk and must share the same sublexical chunk and the same stem extensionthe same stem extension
Is this the case?Is this the case?
What are the major schemas?What are the major schemas?
Are they all local?Are they all local?
Separate networks for –a Separate networks for –a and –i verbsand –i verbs
kam
kap
kak
kaz
kar
kaj
kad
kim
xam
kum
kajm
kach
-a-i
The most connected –a verbsThe most connected –a verbsmin number of neighbors = 20min number of neighbors = 20
The most connected –i verbsThe most connected –i verbsmin number of neighbors = 35min number of neighbors = 35
Adding some less connected –i Adding some less connected –i verbsverbs
(min #of neighbors = 20)(min #of neighbors = 20)
ConclusionConclusion
There are large clusters of verbs in the There are large clusters of verbs in the lexicon in which all verbs are similar to lexicon in which all verbs are similar to each other in exactly the same way, which each other in exactly the same way, which could give rise to schema formation.could give rise to schema formation.
Many of such schemas would not involve Many of such schemas would not involve sharing segments that are adjacent to the sharing segments that are adjacent to the suffix.suffix.