Class 1: Introduction and OT Basics Adam Albright (albright@mit.edu) LSA 2017 Phonology University of Kentucky
Mechanics ▶ Syllabus ▶ Office hours ▶ Background ▶ Class website: lsa2017.phonology.party Introduction Constraints Ranking Modeling distributions Practice References 1/42
What is the goal of phonological analysis? ▶ Describing corpora ▶ Describing lexicons ▶ Describing speakers Introduction Constraints Ranking Modeling distributions Practice References 2/42
What is the goal of phonological analysis? ▶ Describing corpora ▶ Describing lexicons ▶ Describing speakers Introduction Constraints Ranking Modeling distributions Practice References 2/42
Phonology as a function We seek to model the function that speakers use to assign probability distributions over surface (output) representations ▶ In general (unconditioned): possible/probable vs. impossible/improbably morphemes, words,etc. ▶ What is a possible word/surface form ▶ Conditioned: the morpheme /ætam/ is pronounced [æɾəm] when no overt suffix is added, not *[ətʰam]. ▶ But the morpheme /ɔtʌm/ is pronounced [ɔɾəm], not *[æɾəm] ▶ What is a possible output for a given word ▶ This function is the grammar ▶ We indirectly observe the distribution that it assigns (pronunciations, acceptability judgments, etc.), infer the function ▶ Language learners have even less evidence, and yet they converge on similar functions Introduction Constraints Ranking Modeling distributions Practice References 3/42
Frequency vs. grammaticality ▶ We seek to model what speakers actually know about distributions ▶ Just because we can observe a restriction in a wordlist, no guarantee that speakers encode it in precisely this form (or at all) ▶ Sanity check: generalization Introduction Constraints Ranking Modeling distributions Practice References 4/42
Generalization: new words ▶ If a sequence is illegal, it will be avoided in new words, e.g., coined or borrowed ▶ English: Acronyms/initialisms create many #Cl items ▶ PLoS (Public Library of Science), vlog ( v(ideo) log ) ▶ Clippings sometimes do as well: (we)blog ▶ However, #tl, #dl are never generated Introduction Constraints Ranking Modeling distributions Practice References 5/42
Generalization: acceptability judgments Halle (1978) ‘Knowledge unlearned and untaught’ ▶ Which of the following would be possible English words? ▶ ptak , thole , hlad , plast , sram , mgla , vlas , flitch , dnom , rtut ▶ Native English speakers tend to agree that… ▶ Some would be perfectly fine English words: plast , flitch , thole ▶ Some are completely impossible: ptak , hlad , mgla , dnom , rtut ▶ Some are in between: vlas , sram ▶ Generally mirrors attestation of clusters ▶ Attested: #pl, #fl, #θ ▶ Marginally attested: #vl ▶ Unattested: #pt, #hl, #rt, #mgl ▶ ‘Blick’ test: confirms speakers generalize certain facts about phonotactic distributions Introduction Constraints Ranking Modeling distributions Practice References 6/42
Underlearning ▶ English has no words ending in [ɛsp] 1 ▶ Apparently not avoided when the result of truncation ▶ OED: resp(ectable) , Thesp(ian) ▶ Urban dictionary: desp(ondent) ▶ Acronyms/initialisms ▶ DESP: Disability & Educational Support Program, Department of Environmental Science and Policy, Division of Extramural Science Programs, Deployment Extension Stabilization Pay ▶ DJ Devin Skylar Post → [dɛsp] 2 ▶ T ypicality judgments (1 = very non-typical, 9 = very typical) Bailey and Hahn (2001) dɹɛsp 4.67 dɹɪsp 4.58 dɹʌsp 4.04 gɹɛsp 6.17 gɹʌsp 5.54 kɹɛsp 5.67 kɹʌsp 4.96 ɹɛsp 5.13 ɹʌsp 5.46 ʃɹɛsp 2.79 ʃɹɪsp 2.33 tɹɛsp 4.33 tɹʌsp 5.04 1 The OED lists ‘ (the) resp ’, a Lincolnshire dialect word from the 18th and 19th centuries referring to a sheep disease caused by brassica poisoning. 2 http://www.soundclick.com/bands/default.cfm?bandID=340811 Introduction Constraints Ranking Modeling distributions Practice References 7/42
Underlearning ▶ (Colloquial) English lacks words beginning with #skl p t k l ✓ spl — — r ✓ spɹ ✓ stɹ ✓ skɹ ▶ Nonetheless acceptable? ▶ Blick test: [sklæb] ▶ Learned words: sclerosis , sclerenchyma ▶ Borrowings: Sklar , Sklodowski , Skluzacek ▶ Clements and Keyser (1983): an accidental gap ▶ Unattested in the language, but permitted by the grammar Introduction Constraints Ranking Modeling distributions Practice References 8/42
Accounting for such discrepancies ▶ These gaps arguably bump up against a limitation on complexity or nature of phonological restrictions ▶ Final ɪsp# and æsp# both attested; *ɛsp# restriction must specifically target ɛsp# ▶ Clements and Keyser: C 1 C 2 C 3 is tolerated if C 1 C 2 and C 2 C 3 are both tolerated ▶ More generally: grammatical formalism determines which facts can be encoded Introduction Constraints Ranking Modeling distributions Practice References 9/42
Overlearning ▶ Many #CC clusters are unattested in data to ordinary learners ▶ #pt, #lb, #zʒ, #hɹ, #vl, #mg, #jw, #sɹ, #bw ▶ Yet some are judged more acceptable than others ? vl, ? sɹ, ? bw ▶ ▶ *pt, *zʒ, *jw, *mg, … ▶ May reflect prior/innate preferences for some sequences over others ▶ Or, generalization based on properties that go beyond the specific segments involved ▶ E.g., phonological features: fricative + liquid Introduction Constraints Ranking Modeling distributions Practice References 10/42
The upshot ▶ If our goal is to model speakers, we should not assume that grammar includes all observable distributional restrictions ▶ Refined goal: formulate a grammar that distinguishes between sounds/sequences that are accepted by native speakers (‘grammatical’), and ones that aren’t (‘ungrammatical’) ▶ In many cases, a restriction is so robust/abundantly supported in the language that we will take it for granted that the grammar encodes it ▶ English lacks uvular consonants ▶ Japanese lacks word-final stops ▶ Promissory note: must confirm that speakers generalize patterns and restrictions Introduction Constraints Ranking Modeling distributions Practice References 11/42
Encoding restrictions A useful assumption: existing words and new words are generated by the same mechanism ▶ I.e., only difference is that known words have been encountered before ▶ Clearly too strong (exceptions to grammar) ▶ Allows us to make predictions about lexicons/corpora ▶ Allows learners to reverse engineer grammar from lexicon/corpus Introduction Constraints Ranking Modeling distributions Practice References 12/42
Encoding restrictions A baseline: an unrestricted model ▶ With some probability α , draw a previously generated morpheme from the lexicon ▶ Otherwise (probability 1- α ), generate a new morpheme ▶ Randomly draw a segment ▶ With some probability p , stop ▶ Otherwise, repeat ▶ Flat distribution: all sounds contrast in all contexts (no predictability) ▶ Generative models vs. discriminative models Demo: GenerateWords.Unconstrained.pl Introduction Constraints Ranking Modeling distributions Practice References 13/42
The function of phonology We need the grammar to… ▶ Eliminate outputs containing certain sounds ▶ I.e., only certain sounds are allowed ▶ Eliminate outputs containing certain sounds in particular contexts, or particular sequences ▶ I.e., only certain sound combinations are allowed ▶ Or, make these outputs less probable ▶ We’ll ignore gradient distributions for now, and focus on binomial (all or nothing) distributions Introduction Constraints Ranking Modeling distributions Practice References 14/42
But what about transformations? ▶ You might have thought that we need the grammar to change certain sounds to other sounds ▶ This is equivalent to ‘eliminate outputs containing certain sounds, in the context where they are in the input’ ▶ More on this below Introduction Constraints Ranking Modeling distributions Practice References 15/42
Phonological constraints ▶ Constraint-based approaches to phonology provide a convenient and intuitive way to model functions that eliminate particular sounds or strings ▶ Starting simply: allowing some sounds and not others (an inventory of surface phones) ▶ Markedness constraints: specify a configuration that is penalized ( marked ) ▶ Each occurrence in a surface form incurs a violation ▶ Indicator functions: register presence or absence of a given configuration ▶ E.g., *b violated by [tæb], [bɔl], [bɪb] (twice), etc. ▶ Satisfied by [tæp], [kɔl], etc. Introduction Constraints Ranking Modeling distributions Practice References 16/42
Filtering outputs ▶ Constraints act as a filter on outputs ▶ Outputs with fewer violations are better ( more harmonic ) than outputs with more violations ▶ The output(s) with the fewest violations are optimal *b ✓ pa ✓ ✓ da ✓ * ba * Introduction Constraints Ranking Modeling distributions Practice References 17/42
Recommend
More recommend
