West Coast Conference on Formal Linguistics (WCCFL) (38th : 2020)

Synchronic knowledge of phonetically unnatural classes Gallagher, Gillian 2020-03-07

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.

Item Metadata

Download

Media
73804-Gallagher_G_Synchronic_WCCFL38_2020.pdf [ 1.17MB ]
Metadata
JSON: 73804-1.0389852.json
JSON-LD: 73804-1.0389852-ld.json
RDF/XML (Pretty): 73804-1.0389852-rdf.xml
RDF/JSON: 73804-1.0389852-rdf.json
Turtle: 73804-1.0389852-turtle.txt
N-Triples: 73804-1.0389852-rdf-ntriples.txt
Original Record: 73804-1.0389852-source.json
Full Text
73804-1.0389852-fulltext.txt
Citation
73804-1.0389852.ris

Full Text

Synchronic	knowledge	of	phonetically	unnatural	classesGILLIAN	GALLAGHER,	NEW	YORK	UNIVERSITYMARCH	7	2020Phonological	classesPhonologists	study	the	patterning	of	sound	classes	in	language◦ ‘voiced	stops	don’t	occur	word	finally’1Phonological	classesPhonologists	study	the	patterning	of	sound	classes	in	language◦ ‘voiced	stops	don’t	occur	word	finally’We’re	interested	in	the	distribution	of	groups	of	sounds,	not	individual	sounds◦ [ptk]	as	a	class,	compared	to	[bdg]	as	a	classRestrictions	on	individual	sounds	are	often	considered	‘accidental	gaps’◦ They	may	be	mentally	represented,	but	they	are	outside	of	the	core	phonological	system2Phonological	classesPhonological	classes	are	defined	by	features,	which	refer	to	phonetic	properties◦ The	phonetic	definition	may	be	complex	and	fairly	abstractThe	feature	[voice]	is	a	label	for	a	collection	of	language	dependent	properties	that	distinguish	/b	d	g/	from	/p	t	k/◦ vibration	during	closure,	voice	onset	time,	F0	&	F1	at	transition,	vowel	length,	burst	amplitude,	closure	duration,	tongue	root	position,	glottal	height,	etc.In	the	‘normal’	case,	sound	classes	are	both	phonetically	and	phonologically	supported◦ The	classes	that	emerge	from	phonetic	analysis	are	those	that	are	referenced	by	the	phonology3Phonological	classesPhonetics-phonology	isomorphy◦ Phonological	classes	emerge	from	phonetic	analysis	(bottom	up	information)◦ Phonological	classes	are	used	in	stating	abstract	patterns	(top	down	information)But:	substantial	phonetic	variability	in	the	production	of	a	category	or	class	is	also	the	normToday:	phonetics-phonology	mismatch◦ In	Bolivian	Quechua,	/q/	patterns	as	a	stop	but	is	frequently	lenited [ʁ]	(among	other	variants)Structure	of	the	talk◦ Phonological	description	of	Quechua◦ Evidence	that	speakers	have	learned	the	distribution	of	/q/◦ Phonetic	study	of	/q/◦ Some	problems	for	learning	phonetically	diverse	classes4Quechua5QuechuaQuechua	is	a	language	family	with	∼10	million	speakers	throughout	the	Andes	(Colombia,	Ecuador,	Peru,	Bolivia,	Chile,	Argentina)South	Bolivian	Quechua:	spoken	by	~1.5	million	in	the	central	valleys	and	altiplano of	Bolivia6QuechuaThe	consonantal	inventory	of	Quechua	includes	lots	of	stopslabial dental postalveolar velar uvular glottalplain p t tʃ kejective p’ t’ tʃ’ k’ q’aspirate ph th tʃh kh qhfricative s ʃ x hnasal m n ɲapproximant w l	r ʎ	j7QuechuaThe	consonantal	inventory	of	Quechua	includes	lots	of	stopslabial dental postalveolar velar uvular glottalplain p t tʃ k /q/	=	[ʁ]ejective p’ t’ tʃ’ k’ q’aspirate ph th tʃh kh qhfricative s ʃ x hnasal m n ɲapproximant w l	r ʎ	j8[ʁ]	is	a	stop:	syllable	structureStops	can’t	occur	before	other	consonants,	or	word-finally.misk'i 'delicious' t'anta 'bread' waʎpa 'chicken’ʎaxta 'town' pampa 'plain' miɾkha 'freckle’*mipk’i,	*watʎa etc.Like	the	other	stops,	[ʁ]	cannot	occur	in	pre-consonantal	or	final	position.ʁaymi ‘festival’ *yaʁmip’isʁo ‘bird’ *siʁp’o9[ʁ]	is	a	stop:	cooccurrence restrictionsEjectives	and	aspirates	occur	word	medially	in	words	with	initial	fricatives	or	sonorantsɾit'i 'snow' mosqhoj 'to	dream’saʧ'a 'tree' ʎimphi 'color'Ejectives	and	aspirates	may	not	follow	other	stops	in	the	word*tant'a *posqhoj*kaʧ'a *ʧimphi[ʁ]	patterns	with	the	stops	– it	cannot	be	followed	by	ejectives	or	aspirates	later	in	the	word*ʁap'a *ʁatha10Behavioral	evidence	for	the	status	of	[ʁ]11Evidence	for	phonological	classesDistributional	patterns	(syllable	structure,	cooccurrence	restrictions)	and	inventory	structure	constitute	phonological	evidence for	a	class	of	stops	that	includes	[ʁ]	(=	/q/)We	can	also	look	for	behavioral	evidence that	speakers	make	use	of	such	a	representationGeneral	methodology:	compare	speakers’	treatment	of	nonsense	words	◦ control	forms:	phonotactically legal	medial	ejective:	[map’i]◦ unattested	combinations	with	true	phonetic	stop:	*[kap’i]	◦ unattested	combinations	with	sonorant	uvular	‘stop’:	*[ʁap’i]Two	types	of	tasks◦ Repetition:	listen	to	a	word	and	repeat	it◦ Forced	choice	wordlikeness judgment:	listen	to	two	words,	pick	which	one	sounds	more	‘natural’12Repetition:	methods19	balanced	Quechua-Spanish	bilinguals	in	Cochabamba,	BoliviaInstructions:	listen	to	a	word	and	repeat	exactly	what	you	hearcontroln	=	10velar-ejectiven =	10uvular-ejectiven =	10fillern	=	30maʎʧ'isit'uwajk'i*kaʎʧ'i*kit'u*kap'a*ʁaʎʧ'i*ʁet'u*ʁap'aʧaxriʎaskusantu13Repetition:	resultsAccuracy	differs	between	the	three	categories◦ control	>	uvular-ejective	(p <	0.001)◦ more	errors	on	*[ʁap’i]	than	on	grammatical	[map’i]◦ uvular-ejective	>	velar-ejective	(p <	0.01)◦ more	errors	on	*[kap’i]	than	on	*[ʁap’i]140255075100control uvular-ejective velar-ejectivepercent correctcontrol ʁ-ej k-ejpercent	correctForced	choice:	methods26	balanced	Quechua-Spanish	bilinguals	in	Cochabamba,	BoliviaInstructions:	listen	to	a	pair	of	nonsense	words,	choose	which	one	is	more	‘natural’All	pairs	contrasted	the	positions	of	C1 and	C2 in	C1V(C)C2V	strings◦ Control	– what	is	the	general	preference	for	ejectives	in	initial	or	medial	position?◦ Velar-ejective/uvular-ejective	– how	does	the	positional	preference	change	when	medial	is	unattested?controln	=	12velar-ejectiven =	12uvular-ejectiven =	12fillern	=	24maʎʧ'i - ʧ'aʎmisiɲʧ'u - ʧ'insuwajʧ'i - ʧ'ajwa*kaʎʧ'i - ʧ'aʎki*kiɲʧ'u - ʧ'iŋku*kap'a – p'aka*ʁaʎʧ'i - ʧ'aʎʁe*ʁeɲʧ'u - ʧ'eɴʁo*ʁap'a – p'aʁaʧaxri - raxʧiʎasku – kasʎusantu - tansu15Forced	choice:	resultscontrol:	prefer	medial	ejective◦ 57%,	e.g.	maʎʧ’i >	ʧ’aʎmiʁ-ej:	disprefer medial	ejective◦ 43%,	e.g.	*ʁaʎʧ'i >	ʧ’aʎʁek-ej:	disprefer medial	ejective◦ 40%,	e.g.	*kaʎʧ'i >	ʧ’aʎki◦ control	>	uvular-ejective,	(p =	0.007)◦ uvular-ejective	≈	velar-ejective,	(p =	0.60)16255075control uvular velarchose medialcontrol ʁ-ej k-ejchose	medialSummary	&	interpretationIn	both	tasks,	control	stimuli	differ	from	k-ejective	and	[ʁ]-ejective◦ k-ejective	and	[ʁ]-ejective	combinations	are	systematically absent◦ the	grammar	penalizes	forms	with	these	combinations*k-ejective	has	stronger	effect	than	*[ʁ]-ejective	in	repetition;	no	difference	in	forced	choice◦ Could	be	a	grammatical	difference,	where	[ʁ]-ejective	forms	are	penalized	less◦ Or	[ʁ]-ejective	forms	are	equally	ungrammatical,	but	pose	fewer	difficulties	in	repetition	taskThere	are	two	grammatical	possibilities◦ Option	1:	Two	separate	restrictions,	*stop-ejective	and	*[ʁ]-ejective	◦ Option	2:	One	grammatical	restriction,	*stop-ejective	([ʁ]	is	a	’stop’,	mapped	to	/q/)	17The	phonetics	of	stops	and	/q/18Phonetic	classesPhonological	classes	of	sounds	can	often	be	defined	based	on	shared	phonetic	properties◦ Often	this	phonetic	definition	is	multi-dimensional◦ It	always	requires	some	analysis	– all	sounds	are	both	similar	and	different	from	all	other	soundsUnambiguous	stops	in	Quechua	have	a	silent	closure,	and	a	burst	of	some	kind◦ The	quality	and	length	of	the	burst	differ	greatly,	by	place	of	articulation	and	laryngeal	category/q/	has	many	different	realizations	(at	least	[q	χ ʁ ɢ]),	many	of	which	lack	either	a	silent	closure	or	a	burst19Some	Quechua	stops[k] [k’] [kh][p] [p’] [ph]20Acoustics	of	/q/[q] [χ] [ʁ][ɢ] closure,	no	burst21Acoustics	of	/q/Spontaneous	speech	collected	via	interview	with	10	bilingual	speakers◦ Voiceless	stop	productions	are	attested,	but	lenition	is	more	common22q no	burst χ ʁ ɢ nall	contexts 25% 9% 13% 45% 19% 582/q/	by	context	– running	speechVariation	reflects	familiar	contextual	preferences	for	voicing	and	manner.23q no burst χ ʁ ɢ nV	_	V 24% 6% 14% 55% 1% 375s	_ 34% 29% 30% 7% -- 70r	_ 13% 10% 34% 39% 3% 97n _ 37% -- -- 32% 32% 19#	_ 56% 6% 11% 22% 6% 18/q/	by	context	– running	speechVariation	reflects	familiar	contextual	preferences	for	voicing	and	manner.24q no burst χ ʁ ɢ nV	_	V 24% 6% 14% 55% 1% 375s	_ 34% 29% 30% 7% -- 70r	_ 13% 10% 34% 39% 3% 97n _ 37% -- -- 32% 32% 19#	_ 56% 6% 11% 22% 6% 18/q/	by	context	– running	speechVariation	reflects	familiar	contextual	preferences	for	voicing	and	manner.25q no burst χ ʁ ɢ nV	_	V 24% 6% 14% 55% 1% 375s	_ 34% 29% 30% 7% -- 70r	_ 13% 10% 34% 39% 3% 97n _ 37% -- -- 32% 32% 19#	_ 56% 6% 11% 22% 6% 18/q/	by	context	– running	speechVariation	reflects	familiar	contextual	preferences	for	voicing	and	manner.26q no burst χ ʁ ɢ nV	_	V 24% 6% 14% 55% 1% 375s	_ 34% 29% 30% 7% -- 70r	_ 13% 10% 34% 39% 3% 97n _ 37% -- -- 32% 32% 19#	_ 56% 6% 11% 22% 6% 18/q/	by	context	– running	speechVariation	reflects	familiar	contextual	preferences	for	voicing	and	manner.27q no burst χ ʁ ɢ nV	_	V 24% 6% 14% 55% 1% 375s	_ 34% 29% 30% 7% -- 70r	_ 13% 10% 34% 39% 3% 97n _ 37% -- -- 32% 32% 19#	_ 56% 6% 11% 22% 6% 18/q/	by	context	– running	speechVariation	reflects	familiar	contextual	preferences	for	voicing	and	manner.28q no burst χ ʁ ɢ nV	_	V 24% 6% 14% 55% 1% 375s	_ 34% 29% 30% 7% -- 70r	_ 13% 10% 34% 39% 3% 97n _ 37% -- -- 32% 32% 19#	_ 56% 6% 11% 22% 6% 18Summary	of	acoustics/q/	is	most	often	intervocalic	and	produced	as	[ʁ]There	is	phonetic	evidence	for	/q/	as	a	stop◦ [q]	production	is	attested	in	all	contexts	(for	almost	all	speakers)◦ Some	contexts	favor	stop	production◦ Some	speakers	produce	more	stops	than	others29Learning	phonotacticswith	phonetic	variation30Phonetic	variation	and	learningBehavioral	and	distributional	evidence	suggests	that	/q/-ejective	combinations	are	systematically	unattested.Speakers	may	have	mapped	all	realizations	to	/q/,	and	learned	a	single	constraint	against	stop-ejective	combinations◦ Phonemic	analysis:	how	do	Quechua	speakers	learn	that	/q/	([q	ʁ χ ɢ…])	is	a	single	category?◦ Featural	analysis:	how	do	Quechua	speakers	learn	that	/q/	([q	ʁ χ ɢ…])	is	a	‘stop’?◦ Belongs	to	the	same	natural	class	as	/p	t	ʧ k/,	has	the	features	[-continuant,	-sonorant]Or,	speakers	may	have	ad-hoc	constraints	against	[ʁ]-ejective,	[χ]-ejective,	etc.◦ Often,	natural	classes	are	used	to	distinguish	systematic	from	accidental	gaps,	with	the	goal	of	avoiding	ad	hoc	constraints	of	this	type◦ Is	a	phonotactic	grammar	of	learnable	from	surface	transcriptions?31Learning	simulationsUCLA	Phonotactic Learner	(Hayes	and	Wilson	2008)◦ Uses	statistical	techniques	to	induce	constraints	on	systematically	underattested patterns	in	a	set	of	learning	data◦ Returns	a	weighted	constraint	set	and	assigns	grammaticality	scores	to	nonce	formsLearning	data:	list	of	10k	Quechua	words	from	Ñawpaqman,	a	newspaper◦ Onsets	only	(unrealistic,	but	easier	to	interpret	for	present	purposes)Representations◦ Phonemic	’UR’	features◦ Surface	featuresMatch	results	to	behavioral	data:	control	>	velar-ejective,	uvular-ejective32Phonemic	representation/q/	is	transcribed	as	[q],	given	the	manner	features	[-continuant,	-sonorant]	(voiceless	stop)/qunqajman/	 given	to	learner	as q	q	m/qunqurikuxtijku/	 q	q	r	k	t	k/qunqujman/ q	q	mThe	model	learns	constraint	*[-continuant,	-sonorant][+cg]	(*stop…ejective)The	model	distinguishes	between	systematic	and	accidental	gaps,	makes	no	other	distinctions33scorecontrol 0	(perfect)k-ej -19q-ej -19Surface	features/q/	is	transcribed	as	[q	ʁ χ ɢ]	based	on	results	from	the	interview	data◦ Transcription	is	proportional	by	context,	e.g.,	initial	/q/	is	62%	[q],	22%	[ʁ],	11%	[χ],	6%	[ɢ]/qunqajman/	 given	to	learner	as q	ɢ m/qunqurikuxtijku/	 q	ʁ r	k	t	k/qunqujman/ ʁ q	m◦ Each	variant	is	given	surface	phonetic	features◦ [q]	=	[-continuant,	-sonorant,	-voice]◦ [ʁ]	=	[+continuant,	+sonorant,	+voice]◦ [χ]	=	[+continuant,	-sonorant,	-voice]◦ [ɢ]	=	[-continuant,	-sonorant,	+voice]34Surface	featuresThe	model	learns	◦ *[-sonorant,	-continuant][+cg]	(no	stop…ejective)◦ *[+dorsal][+cg]	(no	[k	k’	kh q	q’	qh ʁ χ ɢ]…[ejective])The	model	distinguishes	control	forms	from	illegal	forms,	but	predicts	substantial	variation	among	illegal	forms35Surface	featuresk-ejective	&	q-ejective	forms	violate	both	constraints;	other	uvulars or	stops	violate	just	onePrediction:	[k]-ejective	should	be	twice	as	bad	as	[ʁ]-ejective◦ Behavioral	results:	[k]-ejective	is	the	same	or	slightly	worse	than	[ʁ]-ejectivePrediction:	[k]-ejective	should	be	twice	as	bad	as	[p]-ejective◦ No	differences	found	looking	back	through	previous	studies36scorecontrol 0	(perfect)k-ejective -40q-ejective -40scoreʁ-ejective -20χ-ejective -20ɢ-ejective -20p-ejective -20SummaryGrammars	learned	from	phonemic	transcriptions	and	surface	transcriptions	differSurface	transcriptions	result	in	more	constraints,	with	overlap,	predicting	too	much	variation	among	unattested	formsTranscribing	[ʁ χ]	doesn’t	prevent	the	model	from	learning	the	restriction	on	these	sounds	◦ While	these	sounds	aren’t	stops,	they	are	uvular◦ The	model	can	learn	the	distribution	of	[ʁ χ]	without	also	learning	other	constraints	on	accidental	gaps◦ The	model	doesn’t	actually	need	to	find	ad	hoc	constraints	on	individual	segments,	it	can	still	learn	general	constraints	on	classes◦ This	is	a	happy	accident	in	Quechua,	might	not	generalize	to	other	languagesPhonotactic	learning	from	surface	transcriptions	may	be	possible,	but	the	current	model	isn’t	perfect37Discussion38Overview	of	findingsPhonetic	variation	and	sound	change	can	easily	lead	to	phonological	patterns	holding	over	phonetically	unnatural	classes	on	the	surface◦ Quechua	/q/	patterns	like	a	stop	but	is	produced	on	the	surface	as	[q	χ ʁ ɢ]Behavioral	experiments	show	that	Quechua	speakers	learn	a	phonotactic	restriction	of	this	type◦ [ʁ]-ejective	forms	are	treated	as	ungrammatical	by	speakersAcoustic	data	shows	that	lenited forms	are	the	majority◦ [q]	realizations	likely	help	learners	establish	a	“stop”	label	for	this	highly	variable	categoryLearning	simulations	lightly	favor	the	representation	of	a	single	stop	category	/q/◦ With	surface	transcriptions,	overlapping	constraints	predict	excessive	variation	among	absent	forms39Outstanding	questionsThe	repetition	task	supports	a	stronger	restriction	on	[k]-ejective	than	[ʁ]-ejective	forms◦ This	could	be	support	for	phonotactic	constraints	over	surface	representations.◦ No	difference	is	found	on	forced	choiceCould	a	learning	algorithm	be	tweaked	to	learn	a	better	grammar	from	surface	transcriptions?◦ And	how	does	surface	transcription	change	the	learning	problem	in	other	languages?	How	would	speakers	learn	a	‘stop’	representation	from	the	highly	variable	phonetic	signal?◦ How	is	[q	χ ʁ ɢ]	mapped	to	/q/?◦ How	is	distributional	information	integrated	into	the	construction	of	phonological	representations?‘Unnatural’	classes	come	in	many	flavors◦ The	Quechua	pattern	is	supported	by	two	phonological	patterns,	as	well	as	true	[q]	productions◦ Is	there	behavioral	evidence	for	unnatural	classes	in	other	languages	with	different	types	of	support?40Thank	you!41Repetition:	resultsAccuracy	differs	between	the	three	categories◦ control	>	uvular-ejective	(p <	0.001)◦ more	errors	on	*[ʁap’i]	than	on	grammatical	[map’i]◦ uvular-ejective	>	velar-ejective	(p <	0.01)◦ more	errors	on	*[kap’i]	than	on	*[ʁap’i]Errors◦ *[kap’i]	à [k’api],	32%◦ *[kap’i]	à [kapi],	18%◦ *[ʁap’i]	à [q’api],	4%◦ *[ʁap’i]	à [ʁapi],	32%420255075100control uvular-ejective velar-ejectivepercent correctcontrol ʁ-ej k-ejpercent	correct/q/	by	context	- isolationVariation	and	lenition	are	both	common	in	running	speech,	for	all	sounds	in	all	languages◦ Is	there	anything	special	about	the	Quechua	case?Word	list	data	is	perhaps	informative	of	representations,	a	‘canonical’	form◦ 9	speakers	produced	5	isolation	words	with	/q/	in	each	context	(45	tokens	per	cell)43q no burst χ ʁ ɢ# _ 62% 2% 13% 18% 4%V _	V 16% 24% 20% 40% --n _ 18% -- -- 22% 60%

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            data-media="{[{embed.selectedMedia}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.73804.1-0389852/manifest

Comment

Related Items