Gunnar Ólafur Hansson and Kevin McMullin. Poster presented at the Workshop on Learning Biases in Natural and Artificial Language Acquisition, LAGB Annual Meeting 2014. September 1-5, 2014 in Oxford, UK.
Biased learning of long-distance assimilation and dissimilation
1. Biased
Learning
of
Long-‐Distance
Assimila:on
and
Dissimila:on
2.
Methodology
Gunnar
Ólafur
Hansson
and
Kevin
McMullin
Experimental
design:
Three
phases
Example
s3muli
Practice: Exposure to six cvcv-Lv stem-suffix pairs in two tenses
Training: 192 triplets with trisyllabic stems and suffixed forms (-li, -ru)
• Suffix liquids can trigger alternation in preceding stem
• Permutations of 3 binary parameters define 8 experimental groups,
differing only in the encountered patterns of stem-suffix interaction:
• Trigger-target interaction: Harmony vs. Dissimilation
• Trigger-target distance: Short-range (cvcvLv-Lv) vs. Medium-range
(cvLvcv-Lv)
• Evidence at other distance level: Faithful non-alternation (rich
stimulus) vs. No relevant stems included (poor stimulus)
• An additional Control group did not hear any stems with liquids
➤
➤
“Past tense” – toke…toke-li; “Future tense” – mebi…mebi-ru
Stimuli (4 different voices) presented over headphones and repeated aloud
Testing: Stem followed by two realizations of suffixed form (2AFC task)
• Choice between harmonic and disharmonic liquid sequence
• 32 trials for stems at each of three trigger-target distances (96 total)
• Short- (cvcvLv), Medium- (cvLvcv), and Long-range (Lvcvcv)
➤
Examples of test trials (harmonic vs. disharmonic choices):
dotile…dotile-li or dotire-li; tukiri…tukiri-ru or tukili-ru (Short-range)
teriti…teliti-ru or teriti-ru; bilegi…bilegi-ru or biregi-ru (Medium-range)
linode…linode-li or rinode-li; renitu…lenitu-li or renitu-li (Long-range)
4.
Summary
Pa6erns
of
learning
and
generaliza3on
Analysis
• Mixed-effects logistic regression model over binary response data
(N = 10,185; log-likelihood = –4,940.5)
• Including Group × Distance interaction greatly increased fit
• Reported p-values and odds ratios (relative to Control group) shown in
graphs were extracted from the fitted model
Future
research
• General reluctance to extend alternations to highly salient word-initial
position (cf. Becker, Nevins & Levine 2012)
• What is the proper characterization of the “transvocalic” relation?
(Syllable-adjacency? Consonant-tier adjacency? Onset-tier adjacency?)
• Can learners discover (or infer) phonotactic assimilation/dissimilation
patterns that involve blocking by intervening segments of certain kinds?
• Computational properties (complexity, learnability) of possible/attested
vs. impossible/unattested patterns (e.g. Heinz 2010, Lai 2012)
cvcvLv-Lv cvLvcv-Lv Lvcvcv-Lv
S-Harm + ! ? ?
S-Harm-M-Faith + ! – ?
M-Harm ? + ! ?
M-Harm-S-Faith – + ! ?
S-Diss + ! ? ?
S-Diss-M-Faith + ! – ?
M-Diss ? + ! ?
M-Diss-S-Faith – + " ?
3.
Results
Short-range harmony groups:
Learning?
* * (p < 0.01) (p < 0.0001)
O.R. = 2.88 O.R. = 12.88
Control S-Harm S-Harm-M-Faith
Test-item type = Short-range (cvcvLv-Lv)
Proportion harmony responses ([l…l] or [r…r])
0.00 0.25 0.50 0.75 1.00
Medium-range harmony groups:
Learning?
* * (p < 0.0001) (p < 0.05)
O.R. = 4.12 O.R. = 2.23
Control M-Harm M-Harm-S-Faith
Test-item type = Medium-range (cvLvcv-Lv)
Proportion harmony responses ([l…l] or [r…r])
0.00 0.25 0.50 0.75 1.00
Short-range harmony groups:
Generalizing OUT to medium-range?
n.s. n.s.
(p ≈ 0.18) (p ≈ 0.13)
O.R. = 1.55 O.R. = 1.65
Control S-Harm S-Harm-M-Faith
Test-item type = Medium-range (cvLvcv-Lv)
Proportion harmony responses ([l…l] or [r…r])
0.00 0.25 0.50 0.75 1.00
Medium-range harmony groups:
Generalizing IN to short-range?
* * (p < 0.0001) (p < 0.01)
O.R. = 3.64 O.R. = 2.76
Control M-Harm M-Harm-S-Faith
Test-item type = Short-range (cvcvLv-Lv)
Proportion harmony responses ([l…l] or [r…r])
0.00 0.25 0.50 0.75 1.00
All harmony groups:
Generalizing OUT to long-range (word-initial position)?
n.s. n.s. * n.s.
(p ≈ 0.37) (p ≈ 0.13) (p < 0.01) (p ≈ 0.11)
O.R. = 1.34 O.R. = 1.64 O.R. = 2.45 O.R. = 1.68
Control S-Harm S-Harm-M-Faith M-Harm M-Harm-S-Faith
Test-item type = Long-range (Lvcvcv-Lv)
Proportion harmony responses ([l…l] or [r…r])
0.00 0.25 0.50 0.75 1.00
References Hansson, Gunnar Ólafur. 2010. Consonant harmony: long-distance interaction in phonology. Berkeley:
University of California Press.
Heinz, Jeffrey. 2010. Learning long-distance phonotactics. Linguistic Inquiry 41(4): 623–661.
Lai, Y. Regine. 2012. Domain specificity in learning phonology. University of Delaware dissertation.
McMullin, Kevin and Gunnar Ólafur Hansson. In press. Locality in long-distance phonotactics: evidence
for modular learning. Proceedings of NELS 44. GLSA Publications, University of Massachusetts.
Rose, Sharon, and Rachel Walker. 2004. A typology of consonant agreement as correspondence. Language
80(4):475–531.
White, James C. 2014. Evidence for a learning bias against saltatory phonological alternations. Cognition
130:96–115.
Becker, Michael, Andrew Nevins and Jonathan Levine. 2012. Asymmetries in generalizing alternations to
and from initial syllables. Language 88(2): 231–268.
Bennett, William. 2013. Dissimilation, consonant harmony, and surface correspondence. Rutgers University
dissertation.
Finley, Sara. 2011. The privileged status of locality in consonant harmony. Journal of Memory and
Language 65:74–83.
Finley, Sara. 2012. Testing the limits of long-distance learning: learning beyond a three-segment window.
Cognitive Science 36:740–756.
Acknowledgements
Workshop
on
Learning
Biases
in
Natural
and
Ar3ficial
Language
Acquisi3on,
LAGB
Annual
Mee3ng,
Oxford,
2014
(Poster downloadable at http://tinyurl.com/HanssonMcMullin-LAGB2014)
This research was supported by SSHRC Insight Grant 435–2013–0455 to Gunnar Ólafur Hansson and a
UBC Faculty of Arts Graduate Research Award to Kevin McMullin. Special thanks to Carla Hudson Kam
and the UBC Language and Learning Lab, as well as to Jeff Heinz, Alexis Black, James Crippen, Ella
Fund-Reznicek and Michael McAuliffe.
Short-range dissimilation groups:
Learning?
Test-item type = Short-range (cvcvLv-Lv)
Proportion disharmony responses ([r…l] or [l…r])
Control S-Diss S-Diss-M-Faith
0.00 0.25 0.50 0.75 1.00
* * (p < 0.0001) (p < 0.0001)
O.R. = 8.54 O.R. = 10.66
Medium-range dissimilation groups:
Learning?
Test-item type = Medium-range (cvLvcv-Lv)
Proportion disharmony responses ([r…l] or [l…r])
Control M-Diss M-Diss-S-Faith
0.00 0.25 0.50 0.75 1.00
* .
(p < 0.001) (p ≈ 0.062)
O.R. = 3.07 O.R. = 1.84
Short-range dissimilation groups:
Generalizing OUT to medium-range?
Test-item type = Medium-range (cvLvcv-Lv)
Proportion disharmony responses ([r…l] or [l…r])
Control S-Diss S-Diss-M-Faith
0.00 0.25 0.50 0.75 1.00
n.s. n.s.
(p ≈ 0.36) (p ≈ 0.26)
O.R. = 1.34 O.R. = 1.45
Medium-range dissimilation groups:
Generalizing IN to short-range?
Test-item type = Short-range (cvcvLv-Lv)
Proportion disharmony responses ([r…l] or [l…r])
Control M-Diss M-Diss-S-Faith
0.00 0.25 0.50 0.75 1.00
* n.s.
(p < 0.0001) (p ≈ 0.95)
O.R. = 3.97 O.R. = 1.02
All dissimilation groups:
Generalizing OUT to long-range (word-initial position)?
Test-item type = Long-range (Lvcvcv-Lv)
Proportion disharmony responses ([r…l] or [l…r])
Control S-Diss S-Diss-M-Faith M-Diss M-Diss-S-Faith
0.00 0.25 0.50 0.75 1.00
n.s. n.s. n.s. n.s.
(p ≈ 0.48) (p ≈ 0.59) (p ≈ 0.27) (p ≈ 0.17)
O.R. = 1.25 O.R. = 1.19 O.R. = 1.43 O.R. = 1.57
EVIDENCE ENCOUNTERED
IN TRAINING DATA
GROUP SHORT-RANGE
(cvcvLv-Lv)
MEDIUM-RANGE
(cvLvcv-Lv)
Control ∅ ∅
S-Harm harmony ∅
S-Harm-M-Faith harmony non-alternation
M-Harm ∅ harmony
M-Harm-S-Faith non-alternation harmony
S-Diss dissimilation ∅
S-Diss-M-Faith dissimilation non-alternation
M-Diss ∅ dissimilation
M-Diss-S-Faith non-alternation dissimilation
1.
Introduc:on
Locality
rela3ons
in
consonant
harmony
Explaining
the
typological
universal
Only two locality types are attested (Rose & Walker 2004; Hansson 2010):
• Transvocalic: interaction in …CVC… only (“syllable-adjacent”?)
• Unbounded: interaction in relevant …C…C… pairs at any distance
Implicational universal: Interaction at some beyond-transvocalic distance
entails interaction in transvocalic contexts (as well as all further distances).
For example, strictly beyond-transvocalic harmony is unattested.
Hypothesis 1: All nonadjacent dependencies originate historically in
transvocalic contexts. The unattested locality patterns are synchronically
possible (and learnable) in principle, but diachronically inaccessible.
Hypothesis 2: The unattested patterns are synchronically disfavoured or
impossible; an inductive bias restricts the hypothesis space available to
learners (and/or the heuristics for navigating this space).
Ar3ficial
language
learning
Dissimila3on
vs.
harmony
Finley (2011, 2012), using poverty-of-stimulus paradigm:
Adult English subjects exposed to sibilant harmony suffix alternation in
medium-range cvCvcv-Cv contexts generalize this to unseen shorter-range
(transvocalic) cvcvCv-Cv and longer-range Cvcvcv-Cv contexts.
Replicated with different design (see §2 below) for sibilant harmony and
liquid harmony (McMullin & Hansson in press; this poster)
Bennett (2013): Dissimilation = avoidance of (similarity-driven) surface
correspondence relation. Predicts typological mismatches for consonant
harmony vs. dissimilation along various dimensions
• Strictly beyond-transvocalic (rather, “beyond-syllable-adjacent”)
dependency should be possible for dissimilation, but not assimilation
• Empirical support for this hypothesis is rather weak (Sundanese?)
Rich-‐s3mulus
vs.
poor-‐s3mulus
training
We extend this line of investigation along two dimensions:
Ø Learning of nonadjacent consonant dissimilation alternations
Ø Training on “rich-stimulus” learning data: overt evidence of
absence of interaction at certain distances
Including evidence of non-interaction allows the training data to instantiate
locality patterns that are unattested (and impossible?)
• Do learners coerce such patterns into their formally simpler, attested
counterparts? (cf. Lai 2012, White 2014)
University
of
Bri3sh
Columbia
Short-range
(cvcvLv stems)
e.g. pokuri
Medium-range
(cvLvcv stems)
e.g. giluko
Harmony pokuli-li…pokuri-ru giluko-li…giruko-ru
Dissimilation pokuri-li…pokuli-ru giruko-li…giluko-ru
Non-alternation pokuri-li…pokuri-ru giluko-li…giluko-ru
No liquids tikemu…tikemu-li…tikemu-ru