Interpreting sonorityprojection experiments the role of phonotactic modeling

Interpreting sonority-projection experiments: the role of phonotactic modeling Bruce Hayes Department of Linguistics UCLA August 19, 2011 Hayes, Interpreting sonority projection

Sonority o o No exact phonetic definition, but it plays a major role in phonological patterning. A typical arrangement of consonants by sonority: glides >> liquids >> nasals >> obstruents August 19, 2011 Hayes, Interpreting sonority projection 2

Sonority sequencing o o o Sonority Sequencing Principle (Sievers 1881; Jespersen 1904; Hooper 1976; Steriade 1982; Selkirk 1984). Sonority preferentially rises through syllable-initial clusters, and falls through syllable-final clusters. Large rises (resp. falls) are better. A pretty good syllable: [pla] A very mediocre syllable: [pta] (tied) A really terrible syllable: [lpa] (reversed) August 19, 2011 Hayes, Interpreting sonority projection 3

The sonority projection effect o o o Ask an English speaker: How good a syllable is [lba]? (horrible sonority violation) How does it compare with [bda]? (merely bad sonority violation) Idea: [lba] is much worse even though the English speaker never heard either one during language acquisition. August 19, 2011 Hayes, Interpreting sonority projection 4

The effect emerges in controlled experiments o o Ø Ø Example: Daland et al. (2011) In the following slide: Horizontal axis: sonority difference of initial cluster, following theory of Clements (1990) Vertical axis: ratings by native speakers (victory percentage, all possible pairwise comparisons of the stimuli) August 19, 2011 Hayes, Interpreting sonority projection 5

Sonority projection in Daland et al. (2011) August 19, 2011 Hayes, Interpreting sonority projection 6

Earlier experimental literature on sonority projection o English: n n n o Korean: n o Pertz and Bever (1975) Albright (2007) Daland et al. (2011) Berent, Steriade, Lennertz, and Vaknin (2007) Berent, Smolensky, Lennertz, and Vaknin-Nusbaum (2009) Berent, Lennertz, Jun, Moreno, and Smolensky (2008) Mandarin: n Ren et al. (2010) August 19, 2011 Hayes, Interpreting sonority projection 7

Three accounts of sonority projection o Universal constraint set (as in Optimality Theory) n o Relative phonetic difficulty in the production of bad -sonority clusters n o Not the whole story: how could such a set be deployed in a grammar that derives sonority projection? See Redford (2008) and work cited there Generalized from the data the child hears n n English has /br/, /pl/, not */rb/, */lp/ — could this be enough to distinguish /bd/ from /lb/? This is the possibility pursued here. August 19, 2011 Hayes, Interpreting sonority projection 8

Research plan o o Strategy: computational modeling, using Hayes and Wilson’s (2008) phonotactic learner Goal: develop grammars that model sonority projection, generalizing from very minimal training data August 19, 2011 Hayes, Interpreting sonority projection 9

Earlier work modeling English o o Daland et al. (2011) use the Hayes/Wilson learner to project sonority, using English learning data. But sonority projection has been shown for languages with much smaller onset inventories than English – is projection possible in such cases? August 19, 2011 Hayes, Interpreting sonority projection 10

Bwa and Ba o o o Bwa is a fictional language whose branching onsets are limited to stop + glide. Ba is a fictional language with no branching onsets at all. Goal: show that sonority projection is possible, without stipulating the Sonority Sequencing Principle a priori. August 19, 2011 Hayes, Interpreting sonority projection 11

Assumptions I: the feature system o From Clements (1990); each feature defines a cutoff on the Sonority Hierarchy. glides liquids nasals obstruents [vocoid] [approximant] + + − − − − [sonorant] + + + − August 19, 2011 Hayes, Interpreting sonority projection 12

Assumptions II: a “UG” of constraints, all “sonority regulating” o A constraint is sonority-regulating if it looks like one of these: o For initial clusters, half the sonority-regulating constraints are “sensible”, half “silly” – both included. August 19, 2011 Hayes, Interpreting sonority projection 13

Sample list of constraints *[−syllabic][−sonorant] *[−syllabic][−approximant] *[−syllabic][−vocoid] *[−syllabic] *[+sonorant][−sonorant] *[+sonorant][−approximant] *[+sonorant][−vocoid] *[+approximant][−syllabic] *[+approximant][−continuant] *[+approximant][−sonorant] August 19, 2011 *[+approximant][−vocoid] *[−consonantal][−sonorant] *[−consonantal][−approximant] *[−consonantal][−vocoid] *[−consonantal][−syllabic] *[−syllabic][+sonorant] *[−syllabic][+approximant] *[−syllabic][−vocoid] *[−sonorant][+sonorant] *[−sonorant][+approximant] Hayes, Interpreting sonority projection 14

Phonemes of Bwa p b f v m w August 19, 2011 t d s z n l r k g a j Hayes, Interpreting sonority projection 15

Features for Bwa o o Consonants: as given earlier Vowels: no sonority features, only [+syllabic] August 19, 2011 Hayes, Interpreting sonority projection 16

Training data: the full vocabulary of Bwa pa ta ka ba da ga fa sa va za ma na la ra ja wa pwa twa kwa bwa dwa gwa pja tja kja bja dja gja August 19, 2011 Hayes, Interpreting sonority projection 17

Weighting the constraints q o o Feed the Hayes/Wilson learner Bwa, with the a priori constraints just given. Learned a grammar — assigning weights to the constraints Software used: http: //www. linguistics. ucla. edu/people/hayes/ Phonotactics/ August 19, 2011 Hayes, Interpreting sonority projection 18

Testing the learned grammar o o Obtain the “penalty scores” it assigns to 16 syllables that embody every possible sonority sequence Arrows show wellformedness differences that should be observed if sonority is projected. August 19, 2011 wwa wra wma wpa rwa rra rma rpa mwa mra mma mpa pwa pra pma Hayes, Interpreting sonority projection ppa 19

Result o o All and only stop + glide clusters perfect. They served as the empirical “kernel” for successful sonority projection. August 19, 2011 Hayes, Interpreting sonority projection 20

The weights of the learned grammar for Bwa o o All of the sensible sonority-regulating constraints got positive weights. All of the silly ones got zero. August 19, 2011 Hayes, Interpreting sonority projection 21

Simulation for the Ba Language Training data: pa ta ka ba da ga fa sa va za ma na la ra ja wa q Vowels assumed to have sonority — same features as glides q Same as before: o q q set of possible constraints training procedure August 19, 2011 Hayes, Interpreting sonority projection 22

Results for Ba o o Again, sonority projection. This time every cluster is penalized, but differentially. August 19, 2011 Hayes, Interpreting sonority projection 23

Diagnosing the simulations o Playing with various constraint sets, I found that: n n o For projection to happen, the constraints be sonority-regulating. If you include constraints with all possible sequences of sonority features, you don’t get projection Why? August 19, 2011 Hayes, Interpreting sonority projection 24

A very simple sonority hierarchy for diagnosis p m w [sonorant] − + + [vocoid] − − + August 19, 2011 Hayes, Interpreting sonority projection 25

The sonority-regulating constraints Sensible *[+vocoid][−sonorant] *[+vocoid][−vocoid] *[+vocoid] C *[+sonorant][−sonorant] *[+sonorant][−vocoid] *[+sonorant] C *C [−sonorant] *C [−vocoid] August 19, 2011 Silly *[−vocoid][+sonorant] *[−vocoid][+vocoid] *[−vocoid] C *[−sonorant][+sonorant] *[−sonorant][+voicoid] *[−sonorant] C *C [+sonorant] *C [+vocoid] Hayes, Interpreting sonority projection 26

The region of possible clusters for Bwa (examples) *ww *wm *wp *mw *mm *mp ✓pw *pm *pp August 19, 2011 Hayes, Interpreting sonority projection 27

What is banned by sensible sonorityregulating constraints? -- “Upward L’” *[+sonorant] C *C [−vocoid] (plus 7 more) *ww *mw ✓pw August 19, 2011 *wm *wp *mm *mp *pm *pp Hayes, Interpreting sonority projection 28

All “silly” constraints ban a legal cluster (total of 9; not all shown) *ww *wm *wp *mw *mm *mp ✓pw *pm *pp August 19, 2011 Hayes, Interpreting sonority projection 29

What happens when the constraints are weighted? o o o All of the silly constraints forbid ✓[pw] – so they get zero weight. Two sensible constraints together, *[+sonorant]C and *C[−vocoid], could do all the work—and maxent does give them the greatest weights. But the system is cautious—Gaussian prior penalizes big weights on individual constraints So, descriptive burden is shared among the other sensible constraints. Basis of sonority projection: the worse the sonority sequencing of the cluster, the more sensible constraints it violates. August 19, 2011 Hayes, Interpreting sonority projection 30

What have we got? o o An explicit grammar that (qualitatively) matches sonority projection intuitions Learning of the grammars weights from very minimal information: the sonority drop across /bw/ (in Bwa), across /ba/ in (Ba) August 19, 2011 Hayes, Interpreting sonority projection 31

But learning still depends on much a priori knowledge o o o Features that regulate sonority Restriction of sonority constraints to “sonorityregulating ones” Nothing said yet about codas, where sonority rises n n The system must be told to look at the syllable peripheries, so it will generalize properly. See full version of this paper. August 19, 2011 Hayes, Interpreting sonority projection 32

Thank you o o Full paper is available in conference proceedings and on line at http: //www. linguistics. ucla. edu/people/hayes Author email: bhayes@humnet. ucla. edu August 19, 2011 Hayes, Interpreting sonority projection 33