Learnability shapes typology: the case of the midpoint pathology∗

Juliet Stanton, MIT – juliets@mit.edu

AMP 2014 – 9/20/14

1 Introduction

• A way to evaluate a proposal in phonological theory: compare its predictions to the attested typology.

• In OT: practice is to look at predictions of a constraint set by exploring its factorial typology.

• Two questions we can ask:

– Does the constraint set undergenerate (does it fail to predict all attested patterns)?

– Does the constraint set overgenerate (does it fail to predict only the attested patterns)?

• Reaction to undergeneration:

– Undergeneration usually points to a problem with the theory.

– We want our theories to be able to account for the full range of linguistic variation.

• Reaction to overgeneration:

– Slightly more complex situation.

– Whether or not overgeneration is perceived as a problem depends on the analyst’s intuition.

The interest of this talk: exploring possible explanations for overgeneration.

• Potential reasons why a predicted pattern is unattested (beyond ‘we just haven’t found it yet’):

– Reason 1 : the language is not part of thelearner’s hypothesis space, i.e. our theoryof Con is wrong.

– Reason 2 : the learner does not receiveenough exposure to the language, andlearns something else instead.

1 Grammar

2

Grammar’

Learner

• Part of evaluating a factorial typology: determining whether 1 or 2 is responsible for the absenceof a predicted but unattested system.

– If 1 is responsible: Con is wrong, and we need to revise it.

– If 2 is responsible: no modification to Con is necessary.

• In order to know for sure that 1 is responsible, we need to rule out 2 .

∗My thanks to Francesca Cicileo for help with word-counting, and to Adam Albright, Donca Steriade, and participants atMIT’s Ling-Lunch for helpful feedback.

1

1.1 Our focus

• We’ll focus on a certain type of unattested stress system (we will call them midpoint systems; cf.Kager 2012, also Eisner 1997, Hyde 2012), where stress is drawn to the middle of mid-length words.

(1) a. oo

b. ooo

c. oooo

d. ooooo

e. oooooo

f. ooooooo

• Three characteristics of this type of system:

– Stress falls at or close to an edge in short words.

– Stress is drawn to the midpoint in mid-length words.

– Stress falls at or close to an edge in long words.

• What kind of grammar gives rise to such a system?

– Kager (2012): two opposite-edge contextual lapse constraints dominate all others.

– Analysis of (1): *ExtLapseL >> *ExtLapseR >> AlignL

– Definitions:

◦ *ExtLapseL: a * if none of the first three syllables are stressed.

◦ *ExtLapseR: a * if none of the final three syllables are stressed.

◦ AlignL: a * for each syllable separating stress from the left edge.

• In short words: domains of *ExtLapseL and *ExtLapseR overlap fully.

(2) Tableau for trisyllabic words/ooo/ *ExtLapseL *ExtLapseR AlignL

� a. [[ooo]L]Rb. [[ooo]L]R *!

• In mid-length words: the domains of *ExtLapseL and *ExtLapseR only partially overlap.

(3) Tableau for five-syllable words/ooooo/ *ExtLapseL *ExtLapseR AlignL

� a. [oo[o]Loo]R **

b. [oo[o]Loo]R *!

• In long words, the domains of *ExtLapseL and *ExtLapseR don’t overlap at all.

(4) Tableau for seven-syllable words/ooooooo/ *ExtLapseL *ExtLapseR AlignL

� a. [ooo]Lo[ooo]R *

b. [ooo]Lo[ooo]R *! * ***

c. [ooo]Lo[ooo]R *! ****

2

• Various other combinations of varieties of *Lapse and *ExtLapse also give rise to this behavior.

– In (5a): *LapseL >> *LapseR >> AlignL.

– In (5b): *LapseL >> *ExtLapseR >> AlignL.

(5) More examples of midpoint systems

a. *LapseL >> *LapseR >> AlignL

i. [[oo]L]R

ii. [o[o]Lo]R

iii. [oo]L[oo]R

iv. [oo]Lo[oo]R

v. [oo]Loo[oo]R

vi. [oo]Looo[oo]R

b. *LapseL >> *ExtLapseR >> AlignL

i. [[oo]L]R

ii. [[oo]Lo]R

iii. [o[o]Loo]R

iv. [oo]L[ooo]R

v. [oo]Lo[ooo]R

vi. [oo]Loo[ooo]R

1.2 Explaining the gap

• Fact: as far as we know, midpoint systems do not exist.

• Kager (2012) claims Reason 1 is responsible: midpoint systems not part of the hypothesis space.

– How Kager (2012) changes the theory: revises Con and Gen.

◦ Con: removal of contextual lapse constraints (e.g. *ExtLapseL).

◦ Gen: adoption of weakly layered feet (cf. Martınez-Paricio 2013).

– Implications for foot-free theories of stress:

◦ In foot-free theories, contextual lapse constraints are vital for an analysis of stress windows.

◦ Removal of contextual lapse constraints = huge problem for foot-free theories of stress.

• Alternative: midpoint systems are part of the hypothesis space; reason 2 is responsible.

– Possible contributing factors:

◦ Midpoint systems should be fairly rare in the first place.

◦ Some midpoint systems hard to learn: data necessary to learn them are rare.

◦ Learners biased against systems where stress placement depends on word length.

– If this alternative succeeds, no need to exclude midpoint systems from the hypothesis space.

– Implications for metrical theory:

◦ No need to eliminate contextual lapse constraints from Con.

◦ No need to adopt weakly layered feet (though cf. Martınez-Paricio 2013, Martınez-Paricio& Kager 2013 for other arguments).

◦ Midpoint systems no longer a problem for foot-free theories of stress.

• This talk: an exploration of this alternative. Does it work? If not, where and why does it fail?

2 How frequent do we expect midpoint systems to be?

• Plausibility of the alternative depends in part on expected frequency of midpoint systems.

– More plausible if we expect midpoint systems to be rare.

– Less plausible if we expect them to make up 90% of all languages.

3

• How can we estimate the expected frequency of a pattern?

– Null assumption: each ranking of constraints is equally probable.

– In many cases: a single surface pattern corresponds to more than one ranking.

– Expectation: more rankings consistent with a surface pattern, more frequent the surface pattern.

◦ Bane & Riggle (2008): statistically significant link between ranking volume (or r-volume)and the frequency of QI stress systems, assuming Gordon’s (2002) QI constraint set.

◦ R-volume of a pattern = # of total rankings consistent with it.

◦ Higher r-volume = higher frequency.

• What is the r-volume of midpoint systems?

– Kager (2012): systems arise when two opposite-edge contextual lapse constraints are at the top.

– 322,560 (8.89%) rankings of Kager’s constraint set fit this description1.

◦ Kager’s precondition: not sufficient to characterize when midpoint systems arise.

◦ Also necessary: same-edge align ranked high (1), repeated in (6).

(6) *ExtLapseL >> *ExtLapseR >> AlignL

a. [[oo]L]R

b. [[ooo]L]R

c. [o[oo]Lo]R

d. [oo[o]Loo]R

e. [ooo]L[ooo]R

f. [ooo]Lo[ooo]R

◦ If opposite-edge align ranked high: attested one-stress system.

(7) *ExtLapseR >> *ExtLapseL >> AlignL

a. [[oo]L]R

b. [[ooo]L]R

c. [o[oo]Lo]R

d. [oo[o]Loo]R

e. [ooo]L[ooo]R

f. [ooo]Lo[ooo]R

◦ 166,080 (4.58%) rankings of Kager’s constraint set give rise to midpoint systems.

• How many midpoint systems do we expect to see?

– Midpoint systems: 4.58% of all systems with one stress.

– Systems with one stress = 75.75% of Gordon’s (2002) survey. I assume this is representative.

– Joint probability: 3.47% of all languages should be midpoint systems.

• As of August 2014, 510 languages represented in StressTyp (Goedemans & van der Hulst 2009).

– Expected: 3.47% of these (or 18) should be midpoint systems.

– Observed: none of them are midpoint systems.

– 18 is not a ton, but the difference between 0 and 18/510 is significant.

1R-volumes here and elsewhere are calculated by hand.

4

3 Extended midpoint systems pose a learnability problem

• Let’s take a closer look at the midpoint system we saw in the introduction.

(8) *ExtLapseL >> *ExtLapseR >> AlignL

a. [[oo]L]R

b. [[ooo]L]R

c. [o[oo]Lo]R

d. [oo[o]Loo]R

e. [ooo]L[ooo]R

f. [ooo]Lo[ooo]R

• Note: if we ignore long words (8e-f), this is just antepenultimate stress!

– When the domains of *ExtLapseL and *ExtLapseR overlap (8a-d), both will be satisfied.

– Relative ranking of *ExtLapseL and *ExtLapseR only visible where they conflict.

• Implication: a learner has to see 6+ syllable words to learn the pattern. Pretty widely assumedthat long words are rare (and I’ll show this); perhaps (8) would be hard to learn.

• Very important: long-word idea only works for a subset of midpoint systems.

– Long-word difficulty: arises when both contextual lapse constraints are varieties of *ExtLapse.

◦ *ExtLapseL, *ExtLapseR: trisyllabic window.

◦ Conflict not apparent until six-syllable words, when they cease to overlap ([ooo]L[ooo]R).

– Long-word difficulty: does not arise when one is a context-sensitive variety of *Lapse.

◦ *LapseL, *LapseR: disyllabic window.

◦ Conflict apparent in shorter words – windows are shorter (e.g. [oo]L[oo]R, [oo]L[ooo]R)

• For now: we will focus on the subset (≈25%) of midpoint systems where the contextual lapseconstraints are *ExtLapseL and *ExtLapseR (we’ll call these extended midpoint systems).

Claim: extended midpoint systems are hard to learn. This is why they’re unattested.

3.1 Long words are rare

• Known: in lexica, short words outnumber long words (e.g. Piantadosi et al. 2011).

• New result : in text corpora, short words also outnumber long words.

– Studied: distribution of word lengths in Mark (New Testament), for 102 languages2.

– Result: a lot of variation, but cross-linguistically, long words are rare.

• For the ‘average’ learner: words containing 6+ vowels make up 1% of the input.

– For (8): average learner receives evidence that *ExtLapseL >> *ExtLapseR in 1% of forms.

– Patterns dependent on 6+ vowel words, like (8), might be difficult to learn.

2I am grateful to Francesca Cicileo for her help with this part of the project.

5

Figure 1: Word length distributions, with median overlay

# of vowels 1 2 3 4 5 6 7Median 34.97% 31.70% 18.30% 9.00% 2.60% 1.00% 0.00%Range <1%-65% <1%-47% <1%-33% <1%-24% <1%-21% 0%-15% 0%-10%

3.2 Modeling extended midpoint systems

• The question: If we give a learner realistic information about word length distribution, does it havea hard time learning extended midpoint systems?

• How address this question:

– Create a learner informed by results of the word length study.

– Compare the learner’s performance on extended midpoint systems to its performance on super-ficially similar, but attested, systems.

• We will focus on a small test set of QI languages: two attested, one midpoint.

(9) Systems in the test set

a. InitialAlignL >> all

i. oo

ii. ooo

iii. oooo

iv. ooooo

v. oooooo

vi. ooooooo

b. Antepenultimate (AP)*ExtLapseR >> AlignL

i. oo

ii. ooo

iii. oooo

iv. ooooo

v. oooooo

vi. ooooooo

c. QI Mid*ExtLapseL >> *ExtLapseR >> AlignL

i. [[oo]L]R

ii. [[ooo]L]R

iii. [o[oo]Lo]R

iv. [oo[o]Loo]R

v. [ooo]L[ooo]R

vi. [ooo]Lo[ooo]R

• Expectation: performance on extended midpoint system depends on the availability of long words.

– If many long words (uncommon situation): learner should find all systems easy to learn.

– If few long words (usual situation): learner should be fine with the attested systems, but findthe extended midpoint system difficult.

6

• Taking a step back: why do we expect this?

– Assumed: in the initial state, all M constraints ranked equally.

– Learning a language = learning all crucial rankings associated with that language.

– If evidence for any given crucial ranking is rare, relevant language will be hard to learn.

◦ AP: requires words of 4+ syllables (frequent) to infer that *ExtLapseR >> AlignL.

◦ QI Mid: requires words of 6+ syllables (rare) to infer that *ExtLapseL >> *ExtLapseR.

3.2.1 Preliminaries: input files, and properties of the learner

• Before we discuss the results, first some information about the learner.

Constraint set

• Kager’s (2012: 1479) anti-lapse constraint set (based on Gordon 2002).

– General anti-lapse constraints:

◦ *Lapse: a * for each sequence of two stressless syllables.

◦ *ExtLapse: a * for each sequence of three stressless syllables.

– Contextual anti-lapse constraints:

◦ *LapseL: a * if neither of the initial two syllables are stressed.

◦ *LapseR: a * if neither of the final two syllables are stressed.

◦ *ExtLapseL: a * if none of the initial three syllables are stressed.

◦ *ExtLapseR: a * if none of the final three syllables are stressed.

– Alignment constraints:

◦ AlignL: one * for each syllable separating stress from the left edge.

◦ AlignR: one * for each syllable separating stress from the right edge.

– NonFinality: a * if the final syllable is stressed.

– WSP: a * for each stressless heavy syllable.

• With this constraint set, all three patterns are OT-consistent.

Input-output pairs

• Properties of the input files:

– Words of one through seven syllables.

– Each possible combination of heavy and light syllables.

– Outputs limited to candidates with only one stress.

• Example: three-syllable words.

Input: hhh hho hoh hoo ohh oho ooh ooo

Outputs: Hhh Hho Hoh Hoo Ohh Oho Ooh OoohHh hHo hOh hOo oHh oHo oOh oOohhH hhO hoH hoO ohH ohO ooH ooO

• Assumption: all combinations of h’s and o’s are equally probable. (Ask me later.)

7

Frequency distributions

• Selected: 5 frequency distributions from the survey data.

– Inuktitut and Luganda have a lot of long words. English and H. Creole have few.

– Portuguese is our ‘average’: it is closest to the median.

Language 1V 2V 3V 4V 5V 6V 7V

Inuktitut 1.5% 5.9% 16.5% 23.6% 24.4% 17.5% 10.6%Luganda 22.7% 21.9% 20.8% 17.5% 10.5% 5.0% 1.6%Portuguese 32.6% 35.4% 18.2% 10.0% 3.0% 0.7% 0.1%English 56.6% 28.0% 11.5% 3.0% 0.6% 0.3% <0.1%H. Creole 58.0% 36.1% 5.1% 0.7% <0.1% 0.0% 0.0%

• Expectation:

– As we move from Inuktitut to H. Creole: extended midpoint systems more difficult to learn.

– No comparable difficulty for the attested systems.

Properties of the learner

• Learner: Magri’s (2012) convergent GLA (cf. Boersma & Hayes 2001)

• For each input file (‘language’): 10 runs, maximally 10,000 trials per run.

• Numbers presented in the results: average trials taken to converge for 10 runs.

3.2.2 Results: extended midpoint systems are hard to learn

• As expected: QI Mid is quite difficult to learn with few long words.

Figure 2: Results for QI systems

Learner Initial AP QI MidInuktitut 3 14 32Luganda 3 15 66Portuguese 4 22 351English 9 68 929H. Creole 12 254 10,000+

• A similar result holds for QS systems: extended midpoint systems much harder to learn.

8

3.3 Local summary

• What the modeling results show: extended midpoint systems are much more difficult to learn thansuperficially similar, but attested, systems.

• Main point: learning an extended midpoint system would require a learner to be exposed to a fairnumber of very long words. For most learners, this is not a realistic situation.

Claim: A factor contributing to the absence of midpoint systems: some are difficult to learn.(A learner would only rarely encounter the forms necessary to learn them.)

4 Exploring the predictions: attested long-word phenomena

• Appeal to the infrequency of long words helps explain the absence of extended midpoint systems.

– Claim: long-word dependence is a contributing factor to the absence of some midpoint systems.

– Prediction: all systems dependent on long words should be improbable.

– I’ll refer to this prediction as the long-word hypothesis.

• Question: does the long-word hypothesis make the right predictions?

• This section: two classes of languages where long words appear to matter.

• Findings:

– In some cases: the behavior of long words is predictable from the behavior of short words.

– In others: patterns dependent on long words occur in languages with many long words.

4.1 Binary plus lapse systems

• Two types of binary plus lapse systems:

– Binary plus external lapse: binary alternation with lapse at an edge.

– Binary plus internal lapse: binary alternation with word-internal lapse.

• Examples: Pintupi (Hansen & Hansen 1969) (10a), Garawa (Furby 1974) (10b).

(10) Binary plus lapse systems

a. Pintupi (Hansen & Hansen 1969)

i. oo

ii. ooo

iii. oooo

iv. ooooo

v. oooooo

vi. ooooooo

b. Garawa (Furby 1974)

i. oo

ii. ooo

iii. oooo

iv. ooooo

v. oooooo

vi. ooooooo

• Relevance: location of lapse potentially unclear until 7-syllable words.

– For example: in real Garawa, lapse adjacent to primary (11a).

– Also conceivable: unattested Garawa′, with the lapse between two secondaries (11b).

9

(11) Binary plus lapse systems

a. Garawa (Furby 1974)

i. oo

ii. ooo

iii. oooo

iv. ooooo

v. oooooo

vi. ooooooo

b. Garawa′

i. oo

ii. ooo

iii. oooo

iv. ooooo

v. oooooo

vi. ooooooo

• Kager (2001): two typological asymmetries in the typology of binary plus lapse systems.

– If the lapse is external, it is always right-edge adjacent.

◦ Constraint proposed to encode this asymmetry:

◦ Lapse-at-End: a * for each lapse not adjacent to the right edge of the word.

– If the lapse is internal, it is always primary-adjacent.

◦ Constraint proposed to encode this asymmetry:

◦ Lapse-at-Peak: a * for each lapse not adjacent to the primary stress.

• The important consequence: they rule out Garawa′. Behavior of seven-syllable words is predictable.

– Pintupi:

◦ Five-syllable ooooo shows us that Lapse-at-End >> Lapse-at-Peak.

◦ The seven-syllable word must then have a right-edge lapse.

(12) Tableau for Pintupi ooooooo/ooooooo/ Lapse-at-End Lapse-at-Peak

� a. ooooooo *

b. ooooooo *!

c. ooooooo *! *!

– Garawa:

◦ Five-syllable ooooo shows us that Lapse-at-Peak >> Lapse-at-End.

◦ The seven-syllable word must then have a primary-adjacent lapse.

(13) Tableau for Garawa ooooooo/ooooooo/ Lapse-at-Peak Lapse-at-End

a. ooooooo *!

� b. ooooooo *

c. ooooooo *! *

• Summary: in binary plus lapse systems, behavior of long words is predictable.

– Expectation: binary plus lapse systems should be easy to learn.

(14) Modeling results for binary plus lapse systems3

System Garawa Pintupi

Avg.trials 58 73

3The learner used for the simulations in this section is different in non-crucial ways from the learner described earlier: forone, it needed to be modified to deal with candidates with more than one stress. Feel free to ask.

10

4.2 Ternary stress

• Up next: languages with ternary stress (each stress separated by two syllables: oooooo).

• Relevance: Some ternary systems pose a challenge for the long-word hypothesis.

• Two ternary systems where long words matter:

– (15a): Chugach Alutiiq Yupik (Chugach; Leer 1985)

– (15b): Cayuvava (Key 1961, 1967)

(15) Stress in Chugach and Cayuvava

a. Chugach4

i. ooo

ii. oooo

iii. ooooo

iv. oooooo

v. ooooooo

vi. oooooooo

b. Cayuvava

i. oo

ii. ooo

iii. oooo

iv. ooooo

v. oooooo

vi. ooooooo

• In both cases: learner must to be exposed to long words to learn the correct pattern.

– Chugach: not clear until eight syllables that ternarity is completely general.

– Cayuvava: not clear until six syllables that there is more than one stress per word.

• Important: the predictions of the long-word hypothesis vary by language.

– Languages with few long words: difficulty with long-word phenomena.

– Languages with many long words: anything goes!

• Prediction: if long-word phenomena exist, we will find them in languages with many long words.

Long words in Chugach

• Question: how frequent are long words in Chugach?

– Unfortunately, I don’t know: it has no online Bible translation.

– But in Inuktitut and Inupiatun (also Eskimo-Aleut), they are extremely frequent.

◦ Inuktitut (8+ vowel words): 12.50%

◦ Inupiatun (8+ vowel words): 12.11%

• We don’t know for sure that the word length distribution of Chugach is similar, but another dialectof Yupik, Central Alaskan, certainly has long words (Miyaoka 2012).

(16) angya-cuara-li-yu-kapigte-llru-nric-aaq-sugnarq-aangaboat-small-make-CNNwn-3Rsg.1sg.do-DES-ITS-PST-NEG-but-INF-IND-3sg.1sg.

‘maybe he really did not want to make me a small boat (though he actually made one).’

4Leer reports all stresses as equal; I assume here that the leftmost is primary.

11

• Question: Is the Chugach pattern easy to learn, for a learner that samples long words at the rate anInuktitut learner would encounter them?

• Answer: Yes, very easy. It takes, on average, around 40 trials to converge.

Long words in Cayuvava

• How frequent are long words in Cayuvava? We don’t know.

• But a short text in Key 1967 (72 words) gives us an idea: words of 6+ syllables make up 26.4%.

Syllables 1 2 3 4 5 6 7 8

# of words 5 0 19 21 8 10 5 4% of total 6.9% 0% 26.4% 29.2% 11.1% 13.9% 6.9% 5.6%

• Question: Is the Cayuvava pattern easy to learn, for a learner that samples long words at the ratethey are attested in Key’s short text?

• Answer: Yes, very. It takes, on average, around 40 trials to converge.

4.3 Local Summary

• Aim of this section: test the long-word hypothesis by investigating two classes of stress systemwhere long words appear to matter.

• The investigation yielded two findings:

– Lapses: asymmetries in the typology suggest that the behavior of long words is predictablegiven the behavior of short words.

– Ternary stress: languages with stress patterns where long words matter also appear to havemany long words.

• Conclusion: the long-word hypothesis makes correct predictions about stress typology. We cansafely point at it as a contributing factor to the unattested status of midpoint systems.

5 What about the rest of the midpoint systems?

• Recall from §3: long-word hypothesis helps explains the absence of only extended midpoint systems.

– Still to explain: absence of the other midpoint systems (≈75%).

• This section: points toward a potential way to account for the absence of the remaining 75%.

On counting systems

• Midpoint systems belong to a larger class of systems – counting systems – that are underattested,relative to what we would expect.

– Counting system = system where distance of stress from an edge depends on word length.

– Expected to be fairly frequent: an r-volume of at least 30%.

12

• Counting systems are rare, but attested: I know of three.

– North Kyungsang Korean (NKK) (Kenstowicz & Sohn 2001)

– Kimatuumbi (Odden 1996)

– Icua Tupi (Abrahamson 1968)

(17) Attested counting systems

a. NKK & Kimatuumbi

i. [oo]R or [oo]R

ii. o[oo]R or o[oo]R

iii. oo[oo]R

iv. ooo[oo]R

v. oooo[oo]R

b. Icua Tupi

i. [oo]R

ii. [ooo]R

iii. o[ooo]R

iv. oo[ooo]R

v. ooo[ooo]R

– Generalization: stress moves to the inner edge of the window to avoid *ExtLapse violations.

• Question: why are counting systems so rare?

• Tentative answer: because learners are biased against systems where stress placement depends onword length (cf. also Staubs 2014).

– Young infants rely on stress as a cue for word segmentation (Jusczyk & Houston 1999, others).

◦ Reliable stress system: stress falls at a fixed distance from the edge.

◦ Unreliable stress system: stress placement depends on word length.

◦ If the learner does not know what the words are, she cannot know that stress placementdepends on their length.

– Possible outcome: learner relies on most frequent pattern (cf. Juszcyk et al. 1993) to segment.

◦ Example: midpoint system in (18), with the Portuguese (‘average’) word length distribution.

(18) *LapseL >> *LapseR >> AlignL

a. [[o]L]R Frequency: 33.2%

b. [[oo]L]R Frequency: 35.4%

c. [o[o]Lo]R Frequency: 18.2%

d. [oo]L[oo]R Frequency: 10.0%

e. [oo]Lo[oo]R Frequency: 3.0%

f. [oo]Loo[oo]R Frequency: 0.7%

g. [oo]Looo[oo]R Frequency: 0.1%

◦ 81.8% of words have initial stress. 18.2% have peninitial stress.

◦ If the learner uses initial stress as a segmentation cue, 18.2% of input segmented incorrectly.

◦ In effect: learner acquires wrong lexicon and wrong stress pattern.

• Idea: counting systems inhibit reliable segmentation; learners are biased against them.

• Back to midpoint systems:

– Their absence: symptomatic of a general dispreference for counting systems.

– Why are other kinds of counting systems attested? Higher r-volume (4.58% vs. 30+%).

13

6 Summary and conclusions

• The goal of this talk: to determine whether or not it’s possible to explain the absence of midpointsystems as a consequence of of the following factors:

– Midpoint systems should be fairly rare in the first place.

– Some midpoint systems are hard to learn: data necessary to learn them are rare.

– Learners biased against systems where stress placement depends on word length.

• Conclusions:

– Too early to say these are entirely responsible for the absence of midpoint systems.

– But too early to rule this out, too.

14

References

Abrahamson, Arne. 1968. Contrastive distribution of phoneme classes in Icua Tupi. Anthropological Linguistics10(6). 11–21.

Bane, Max & Jason Riggle. 2008. Three Correlates of the Typological Frequency of Quantity-Insensitive StressSystems. In Proceedings of the tenth meeting of the ACL special interest group on computational morphology andphonology, 29–38. Columbus, Ohio.

Eisner, Jason. 1997. What constraints should OT allow? Paper presented at the Annual Meeting of the LinguisticSociety of America, Chicago. ROA-204.

Furby, Christine. 1974. Garawa phonology Pacific Linguistics, series A. Canberra: Australian National University.

Goedemans, Rob & Harry van der Hulst. 2009. StressTyp: a database for word accentual patterns in the world’slanguages. In The use of databases in cross-linguistics research, 235–282. New York: Mouton de Gruyter.

Hansen, K. C. & L. E. Hansen. 1969. Pintupi phonology. Oceanic Linguistics 8(2). 153–170.

Hyde, Brett. 2012. Alignment constraints. Natural Language and Linguistic Theory 30. 789–836.

Jusczyk, Peter W. & Derek M. Houston. 1999. The beginnings of word segmentation in English-learning infants.Cognitive Psychology 39. 159–207.

Juszcyk, Peter, Angela D. Friederici, Jeannine M. I. Wessels, Vigdis Y. Svenkerud & Marie Jusczyk. 1993. Infants’sensitivity to the sound patterns of native language words. Journal of Memory and Language 32(3). 402–420.

Kager, Rene. 2001. Rhythmic directionality by positional licensing. Paper presented at the Fifth HIL PhonologyConference (HILP 5).

Kager, Rene. 2012. Stress in windows: Language typology and factorial typology. Lingua 122. 1454–1493.

Kenstowicz, Michael & Hyang-Sook Sohn. 2001. Accentual adaptations in North Kyungsang Korean. In MichaelKenstowicz (ed.), Ken hale: a life in language, 239–270. Cambridge: MIT Press.

Key, Harold H. 1961. Phonotactics of Cayuvava. International Journal of American Linguistics 27. 143–150.

Key, Harold H. 1967. Morphology of Cayuvava. The Hague: Mouton.

Leer, Jeff. 1985. Prosody in Alutiiq. In Michael Krauss (ed.), Yupik eskimo prosodic systems: descriptive andcomparative studies, 77–134. Fairbanks, AK: Alaska Native Language Center, University of Alaska.

Martınez-Paricio, Violeta. 2013. An exploration of minimal and maximal metrical feet : University of Tromsø disser-tation.

Martınez-Paricio, Violeta & Rene Kager. 2013. Non-intervention constraints and the binary-to-ternary rhythmiccontinuum. Paper presented at the 21st Manchester Phonology Meeting.

Miyaoka, Osahito. 2012. A grammar of Central Alaskan Yupik (CAY). New York: Mouton.

Odden, David. 1996. Phonology and morphology of Kimatuumbi. Oxford: Oxford University Press.

Piantadosi, Steven T., Harry Tily & Edward Gibson. 2011. Word lengths are optimized for efficient communication.In Paul Kay (ed.), Proceedings of the National Academy of Sciences of the United States of America, vol. 108 9,3526–3529.

Staubs, Robert. 2014. Learning and the position of primary stress. In Robert E. Santana-LaBarge (ed.), Proceedingsof the 31st West Coast Conference on Formal Linguistics, 428–437.

15