Schedule of Talks for Fall 2014
SEPTEMBER 8 -
Myfany Turpin (University of Queensland)
Linguistic fieldwork and song
Jointly hosted by Phorum and FForum
SEPTEMBER 15 -
Megha Sundara (UCLA)
Phonetic similarity biases phonological learning in infants
Researchers have suggested that learners are biased to prefer phonological mappings between sounds that are phonetically similar (Steriade, 2001; Peperkamp et al., 2006; Wilson, 2006; White, 2014). We tested this claim in one artificial language study and one natural language study involving 12-month-old infants. Our results demonstrate that infants generalize phonological learning in ways that are not predicted from the input alone and sometimes even fail to learn patterns available in their input. These findings show that input statistics alone cannot account for how infants learn phonological alternations. Instead, phonological learning is biased by a preference for alternations involving phonetically similar segments.
SEPTEMBER 22 -
Ronald Sprouse and Keith Johnson (UC Berkeley)
The Berkeley Phonetics Machine
We will demonstrate a virtual linux machine that we are calling The Berkeley Phonetics Machine. BPM runs in VirtualBox on Mac, PC, and Linux, and has a number of standard and custom software packages for phonetic research.
Custom Berkeley software:
- ifcformant - Inverse Filter Control formant tracking, which seems to be a very good formant tracking method.
- make_text_grids - An implementation of the Penn Forced Aligner for aligning with marked up transcripts.
- wxklsyn - An implementation of the Klatt speech synthesizer.
- xwaves - A collection of speech signal processing for command-line scripting.
- convertlabel - A tool for converting between audio annotation formats.
- audiolabel - A Python library for reading audio annotation files.
The machine also includes standard packages: opensesame, Praat, Wavesurfer, sox, ffmpeg, edgetrack. The demo will show the machine, and discuss how to download and run it.
SEPTEMBER 29 -
Larry Hyman (UC Berkeley)
Initial Vowel Length in Lulamogi: Cyclity or Globality?
Over the past several decades there has been recurrent skeptism concerning cyclic derivations in phonology, one of the most central tenets of traditional generative and lexical phonology and morphology. Some of the proposed cyclic analyses have been argued not to require cyclicity, or to represent lexical relations that are not totally productive (as in certain cases in English). For those surviving cases, a major strategy within optimality theory has been to capture cyclic relations by surface "output-output" (O/O) constraints. Thus, to take a standard example, génerative and derívative have different stress patterns not because they are literally derived from génerate and deríve, but because the stress of each derivative must agree with the output stress of its respective corresponding base. A particularly explicit (and hence falsifiable) component of O/O correspondence is that "a cyclic Base must be a freely occurring expression, a phrase or a free-standing word (Benua 1997, Kager 1999; Kenstowicz 1996; cf. Bermúdez-Otero 2010, Kiparsky 1998, Trommer 2013 for critical discussion and proposed counterevidence)" (Steriade 2013). In this paper I draw on original data from Lulamogi, a previously almost unstudied Bantu language of Uganda, to show that the most insightful analysis of a vowel length alternation requires either cyclicity or global reference to internal morphological structure and, in many cases, a non-free standing base.
OCTOBER 6 -
Gregory Finley (UC Berkeley)
Recruitment of dual-source binaural cues by hearing impaired listeners
Speaker's disclaimer: This is not a linguistics talk, but it may appeal to those interested in perception, hearing science, or industry research in general.
In this talk I present a project I undertook while an intern at Starkey Hearing Technologies. Listeners with hearing impairment perform simple binaural tasks such as sound localization nearly as well as normal hearing listeners, but for more complex conditions, such as speech recognition in noise, their binaural abilities fall short. How well would these listeners handle a simple task performed on a complex scene? I devised a study to test lateralization of two sources, a male and a female talker, at fine differences in simulated location. Performance suffered somewhat with age and hearing loss, with some older listeners completely unable to perform the task under some conditions. The results' similarities and differences to those of other experimental paradigms are discussed, as are plans for further research.
OCTOBER 13 -
Andréa Davis (U Arizona / UC Berkeley)
When more is not better: Variable input in the formation of robust word representations
A number of studies with infants and with young children suggest that hearing words produced by multiple talkers helps learners to develop more robust word representations (Richtsmeier et al 2009, Rost & McMurray 2009, 2010). Native adult learners, however, do not seem to derive the same benefit from multiple talkers. Two word learning studies, with native English-speaking adults and with second-language English-speaking adults, were conducted. In both studies, participants learned 4 new minimal English-like minimal pair words either from a single talker or from multiple talkers. They were then tested with a) a perceptual task, in which they saw the two pictures corresponding to a minimal pair, heard one of the pair, and had to choose the picture corresponding to the word they heard b) for the native English speakers a speeded production task, in which they had to repeat the words they had just learned as quickly as possible. Unlike infants and young children, native English-speaking adults did not differ significantly between the multiple talker group and the single talker group, either in perceptual accuracy, or in production. However, the second language English speakers did differ significantly, but only when they were less proficient in English. The more proficient speakers, like the native speakers, did not differ. Taken together, these results suggest that proficiency plays a role in whether learners benefit from variable input.
OCTOBER 20 -
David Conant (UCSF)
A multi-modal imaging system for simultaneous measurement of speech articulator kinematics compatible with human electrophysiology
Speech articulation involves the rapid, coordinated movement of speech articulators (e.g. lips, jaw, tongue, and larynx). Most neuroscience investigations of speech have relied upon static, binary features instead of dynamic articulator kinematics. However, a complete neurobiological understanding of speech motor control requires determining the relationship between simultaneously recorded neural activity and the kinematics of all articulators. Many speech articulators are internal to the vocal tract, and so simultaneously tracking the kinematics of all articulators is difficult, especially in the context of human electrophysiology recordings.
Here, we describe a noninvasive, multi-modal imaging system for simultaneously tracking the movement of the lips, jaw, tongue and larynx for articulator tracking in human neuroscience at the bedside. We combined three non-invasive methods previously used separately: videography to track the lips and jaw, electoglotiography to monitor the larynx, and ultrasonography to track the tongue. To characterize this system, we recorded articulator positions and acoustics from six speakers during multiple productions of nine American English vowels.
We first describe processing methods for the robust extraction of kinematic parameters from the raw signals and to alignment/scaling methods to account for artifactual variability across recording conditions. These results generally confirm the importance of tongue height and tongue frontness, and the first and second formants, but show significant cross-subject variability. We used unsupervised matrix factorization techniques to extract 'basis sets' of vocal tract 'shapes' associated with different vowels. This data-driven approach may preserve information about vocal tract 'shape' better than traditional point extraction methods. Finally, we developed a statistical speech synthesizer to convert measurements of the vocal tract to audible speech and were able to reconstruct perceptible speech from measured articulatory features. These results demonstrate a multi-modal experimental system to non-invasively monitor articulator kinematics during speech articulation, and describe novel analytic methods for relating kinematic data to speech acoustics.
OCTOBER 27 -
Martin Kohlberger (Leiden University)
Phonetic evidence for gradience: Spanish resyllabification and voicing reconsidered
In this presentation, I will discuss how subtle phonetic cues can provide crucial insight into the complexity involved in various phonological phenomena. Two experiments will be discussed: one on Spanish resyllabification and one on Spanish voicing in whispered speech.
Word-final consonants in Spanish are commonly assumed to undergo resyllabification across a word boundary before a following vowel, e.g. /los#otros/ ‘the others’ is realised as [lo.s|o.tros]. However, in many dialects of Spanish, word-final prevocalic consonants (‘derived onsets’) pattern phonologically with canonical codas and distinctly from canonical onsets. This property of derived onsets has been the subject of much interest in the phonological literature, and has led some linguists to question whether resyllabification indeed applies in all Spanish dialects. In this presentation, evidence for resyllabification is evaluated based on acoustic data from 11 speakers of Castilian Spanish. The results show that word-final prevocalic /s/ has increased duration compared to coda /s/, but at the same time, it is shorter compared to word-initial /s/. This result is incompatible with a full resyllabification hypothesis, which would predict word-final prevocalic /s/ to be indistinguishable from canonical onsets. An alternative in the form of partial resyllabification is considered, the role of the syllable as a relevant unit in explaining /s/-sandhi in Spanish is further discussed.
An underlying voicing contrast tends to be preserved in whisper, even though phonetic voicing as such is not present. This study investigates whether whispered speech also contains phonetic cues to instances where voicing is not lexically contrastive, but where it arises contextually through assimilation to a following obstruent. Spanish speakers read 8 repetitions of 6 test items where word-final /s/ appeared in the context of a following fortis or lenis obstruent (e.g. gatas tensas, sopas densas). Of the 8 repetitions, 4 involved normal phonation and 4 involved whisper. A number of cues were analysed as potential correlates of sibilant voicing. Results show that the contrast between sibilants followed by fortis and lenis stops was preserved in vowel duration and vowel-consonant intensity differences. The contrast between sibilants followed by fortis and lenis stops in whispered speech indicates that aspects of voicing assimilation may also be found in the absence of vocal fold vibration. This result contributes to our understanding of how different cues are integrated in signalling assimilatory voicing. As the cues observed in whispered speech cannot be secondary to the presence of vocal fold vibration, they should be viewed as independent and speaker- controlled exponents of voicing assimilation.
NOVEMBER 3 -
Larry Hyman (UC Berkeley)
Positional Prominence vs. Word Accent: Is there a difference?
One of the major unresolved issues in the study of word-accentual systems is determining what exactly counts as accent. In languages such as English, where prominent syllables are uncontroversially identified by a combination of effects on both the suprasegmental features of pitch, duration and intensity as well as on segmental realizations, there is no hesitation in attributing these effects to stress and metrical structure. Controversy arises in languages where the effects are less pronounced, have little or no effect on segments, or mark phrasal domains rather than words. A case in point is the common Bantu phenomenon of penultimate vowel lengthening, which, although often identified as accent (cf. Downing 2004), generally characterizes the penultimate syllable of phonological or intonational PHRASES, rather than the word (Hyman 2013). The intonational nature of such alleged “accentual” lengthening is particularly striking in a language like Shekgalagari, where penultimate vowel lengthening occurs in declarative utterances, but not in questions or imperatives (Hyman & Monaka 2011). The task of unambiguously identifying accent is often further complicated in languages with tone or so-called pitch-accent. Given the wide range of potential effects on tone, length, and segmental phonology in prominent vs. non-prominent syllables, there is no guarantee that these effects will exhibit the same “metrical coherence” (Dresher & Lahiri 1991) known from Germanic and other languages which care a lot about marking their stressed syllables. If the prominent positions conflict, the question is whether they are “accentual” in the normal sense.
In an attempt to resolve this question, in this paper I address two Bantu cases where the positional prominence effects are clearly word level and reasonably subject to a metrical (accentual) interpretation, although one that might not completely “cohere”. The first is Punu (Fontaney 1980, Blanchon 1995), spoken in Gabon. In this language greater consonant and vowel contrasts suggest that the initial root syllable is accented, tone suggests the penultimate syllable is accented, and vowel length suggests that both positions are accented. The second case comes from Lulamogi, a small, understudied language spoken in Uganda, which we have been investigating at Berkeley since last August. In this language vowel length suggests that the initial, second, penultimate and final syllables of the stem are accented, while tone suggests it is the penult. (The segmental phonology does not show significant positional prominence effects.) Although there are problems in both cases in accommodating different prominent positions and questions about ultimate foot structure, I will suggest that at least a root-initial trochee is motivated in Punu and a word-final trochee in Lulamogi. I conclude with discussion of one final complication from Luganda: whether it is moraic vs. syllabic positions which are prominent.
NOVEMBER 10 -
Ben Parrell (UC Berkeley (Psychology))
Consonant lenition: prosodic causes, acquisition, and sound change
I will show how lenition, rather than being a strictly categoric or allophonic process, can result from the interaction between invariant production goals (i.e. phonological targets) and prosodic structure. Reduction of coronals stops in English and Spanish will be compared and shown to both arise from such interactions. I will also show a computational model of this phonologic-prosodic interaction, and examine how acquisition of these patterns can be modeled in the same framework, as well as what this model can tell us about the relationship between synchronic lenition patterns and sound change.
NOVEMBER 17 -
Sharon Inkelas (UC Berkeley)
Affix order and informativity: a pilot study of Turkish
Preliminary analysis of a Turkish corpus suggests a link between affix order and affix informativity, i.e. predictability in context. The results of this pilot study show that informativity of a suffix in the context of the preceding morpheme is negatively correlated with distance from the root. This result may help to understand the nature and function of affix ordering templates in highly affixing languages.
DECEMBER 10 -
LSA Practice Talks (Jevon Heath; Clare Sandy)
Note unusual time, day, and location: 3401 Dwinelle at 1:00 PM!
DECEMBER 15 -
Robert Eklund (Linköping University)
The neural correlates of filled pauses: An fMRI study of disfluency perception
A characteristic of spontaneous spoken language is that no one is completely fluent. When speaking (or signing) everyone exhibits a certain degree of disfluency, to use the most common term. The average frequency of disfluency has been reported to be around 6% at word-level (Eklund, 2004; Bortfeld et al., 2001; Brennan & Schober, 2001; Fox Tree, 1995; Oviatt, 1995). This talk presents results from a perception study of disfluency in spontaneous speech (Eklund, 2004:175-196). Unfilled pauses (silences; UPs) and filled pauses (“eh”; FPs) were excerpted from spontaneously produced dialogs and were played to subjects in an fMRI experiment where the subjects acted as silent interlocutors.
The results exhibited increased activity in Primary Auditory Cortex for both types of stimuli. However, and more interestingly, FPs, but not UPs, also elicited significant modulation in the Supplementary Motor Area (Brodmann Area 6). To the best of our knowledge, this study represents the first neurocognitive confirmation of the well-known difference between FPs and all other kinds of speech disfluency. The results also provide an explanation for the previously reported beneficial effect of FPs on reaction times in speech perception, and is also interesting from the alleged role of FPs as “floor-holding devices” in human dialog, since FPs seemingly activate motor programs in the listener.
The observed activation also has implications for various other perspectives of human communicative and motor actions, such as mirror neuron theory, motor theory of speech perception, and even help shed light on the oft-reported very short inter-speaker intervals (ISIs) in human-human dialog.
Schedule of Talks for Spring 2014
JANUARY 27 -
Nik Rolle (with Marine Vuillermet)
Berkeley; Dynamique du Langage, CNRS
Accent Assignment, Clash, and Resolution in Ese Ejja (Takanan)
Ese Ejja is a Takanan language spoken in the Bolivian Amazon. This talk presents a working analysis of the distribution of accent in Ese Ejja phonological words headed by nouns and verbs. Phonological words are (partially) defined as having a primary accent which falls on one of the first three syllables of the word, whose consistent phonetic correlate is high pitch. We argue in this talk that the distribution of primary accent depends on a complex interaction of factors, including (1) inherent accent of a word, (2) accent assignment from affixes/clitics, (3) accent assignment based on major part of speech (noun vs. verb) and word class (e.g. transitive vs. intransitive verb), (4) rules of accent clash resolution, (5) rules of (trochaic) footing, and (6) restrictions on the 3-syllable primary accent window. Of particular theoretical interest will be the proposed rules of accent assignment and accent clash resolution, the lack of alignment between trochaic feet and the 3-syllable accent window, and the correlation between grammatical categories and distinct phonological patterns.
FEBRUARY 3 -
Stephanie Shih
Berkeley / Stanford
The phonological conditioning of morphosyntactic variation
Speakers encounter a number of morphosyntactic choices in language production, giving rise to variation between lexical items, word orders, and constructions, as in the English possessive:
(1) the car’s wheel ~ the wheel of the car
Previous work has shown that morphosyntactic variation of this kind is conditioned by numerous factors, ranging from syntactic and semantic to usage-based, psycholinguistic, and sociolinguistic (e.g., Bresnan et al. 2007; Tagliamonte & Baayen 2012). In some cases, prosodic factors (e.g., prosodic phrasing) have been suggested to affect word order choices as well (e.g., Zec & Inkelas 1990; Zubizarreta 1998), but for the most part, the influence of phonology has largely been discounted in previous work as a contributor to morphosyntactic choice due to theoretical and empirical limitations (e.g., Zwicky & Pullum 1986; though cf. McDonald et al. 1993; a.o.).
In this talk, I argue that lower-level phonological factors, including segmental and metrical well-formedness preferences, are at work in conditioning morphosyntactic choices. Such effects have been heretofore relatively unavailable for study because they involve probabilistic patterns in natural language use, requiring ecologically-valid datasets and sensitive quantitative methodologies. Here, I present the results of three corpus-based studies. The first examines phonological effects on lexical choices in English forename-surname pair formation (e.g., Susan Smith versus Suzanne Smith), drawing on a large-scale corpus of Facebook usernames. The second looks at the influence of phonological factors on spoken English possessive variation (e.g., the car’s wheel ~ the wheel of the car) in the Switchboard Corpus, relative to other known syntactic, semantic, psycholinguistic, and sociolinguistic factors that also affect this choice. The third case study turns to cross-linguistic evidence from a web text corpus of Tagalog adjective-noun word order variation (e.g., magadang baba’e ~ baba’eng maganda ‘beautiful woman’), which involves phonologically-conditioned surface allomorphy that contributes to the choice of word order. The results from the three case studies demonstrate that morphosyntactic choices can optimize for segmental (e.g., phonotactic) and metrical (e.g., rhythmic) well-formedness. While the influence of such low-level phonology is small relative to higher-order semantic and usage-based factors such as animacy and frequency, their influence remains robust across languages as well as across variable phenomena, in lexical, word order, and construction variation. The interaction reported in this talk between phonological information and morphosyntactic choices raises theoretical implications for considering the architecture of the morphosyntax-phonology interface in both formal and psycholinguistic models of language (cf. Zwicky & Pullum 1986; Levelt 1989).
FEBRUARY 24 -
Joe Pater
UMass Amherst
Prof. Pater will be leading us in an informal discussion of the paper linked below. Come prepared with questions!
Pater, Joe, Robert Staubs, Karen Jesney and Brian Smith. 2012. Learning probabilities over underlying representations. In the Proceedings of the Twelfth Meeting of the ACL-SIGMORPHON: Computational Research in Phonetics, Phonology, and Morphology. 62-71.
We show that a class of cases that has been previously studied in terms of learning of abstract phonological underlying representations (URs) can be handled by a learner that chooses URs from a contextually conditioned distribution over observed surface representations. We implement such a learner in a Maximum Entropy version of Optimality Theory, in which UR learning is an instance of semi-supervised learning. Our objective function incorporates a term aimed to ensure generalization, independently required for phonotactic learning in Optimality Theory, and does not have a bias for single URs for morphemes. This learner is successful on a test language provided by Tesar (2006) as a challenge for UR learning. We also provide successful results on learning of a toy case modeled on French vowel alternations, which have also been previously analyzed in terms of abstract URs. This case includes lexically conditioned variation, an aspect of the data that cannot be handled by abstract URs, showing that in this respect our approach is more general.
MARCH 10 -
Marc Ettlinger
VA Northern California Healthcare System
What is artificial language learning?
MARCH 31 -
Pat Keating
UCLA
Phonetics in 1924
This talk reviews several areas of phonetics that were important around 1924, when the Linguistic Society of America was established. The 1920s saw widespread use and expansion of the IPA for phonetic description of languages, including major works on American English phonetics; new speech production laboratories in the U.S. with many specialized instruments; the beginnings of the acoustic theory of speech production; benefits from telecommunications research for basic speech science; and the participation of linguists and non-linguists in phonetics.
APRIL 7 -
Kevin McGowan (with Anna M. Babel)
Stanford; The Ohio State University
Perceiving isn't believing: listeners' expectation and awareness of phonetically-cued social information
There is abundant evidence that listeners are able to use top down social information to alter or enhance the perception of phonetic detail. Extra-linguistic cues to talker identity have been argued to alter (Niedzielski, 1997; Hay & Drager, 2010), guide (Staum Casasanto, 2009), or enhance (Sumner & Samuel, 2007; McGowan, 2011; Szakay et al., 2012) listeners' mapping of available phonetic information to linguistic categories. However, the extent to which listeners are aware of the relationships between social categories and sociolinguistic variables is very much an open question in both sociolinguistics and speech perception. Nevertheless, work in this literature relies on a mixture of early and late tasks that assume a stable link between metalinguistic commentary gathered a relatively long time after the presentation of the stimulus and the listeners' early, online perceptions. We present evidence from a within-subjects matched guise task which suggests that the relationship between percept and late sociolinguistic judgment may be far more complex than has been assumed.
We manipulated whether listeners believed a speaker to be from the Quechua-dominant west of Bolivia or from the Spanish-dominant east. In an unusual twist for this type of experiment, listeners were presented with first one guise and then the other in subsequent blocks with no change in voice. Listeners who were initially told the speaker is Quechua-dominant perform essentially at chance on the listening task and continue to perform at chance on the task when switched to the Spanish-dominant guise. When these same stimuli were presented with the guises reversed, however, clear category boundaries for the vowels in each continuum emerge and are retained across the switch from Spanish to Quechua guise conditions. In contrast to these results, interviews give every indication that participants believed the guise switch had occurred - commenting on everything from the pronunciation to the relative education level of the new speaker. This mismatch between early linguistic judgments and late metalinguistic commentary suggests that while there may be a dual linguistic and social path to speech perception (Sumner et al., 2014), the time course of linguistic perception and social perception requires further careful investigation.
APRIL 21 -
Eric Campbell
UT Austin
Typological perspectives on the tone system of Zenzontepec Chatino
In register tone systems (Pike 1948) with an unspecified or toneless category, a mid tone is typically the unspecified one (Maddieson 1978; Hyman 2012). In this talk I will present the tone system of Zenzontepec Chatino (Otomanguean, Mexico) and argue that there is a three-way tonal contrast on the mora: high tone (H) vs. mid tone (M) vs. no tone (Ø), with the last having a default low or mid-to-low falling pitch. Several types of evidence support treating the lowest category as toneless in Zenzontepec Chatino, the strongest coming from tonal processes in connected speech, which are illustrated with audio examples from naturalistic texts. Then, I discuss (i) why the typology has excluded such a system, which as far as I've seen has only been reported elsewhere once (Paster 2003), (ii) how the system may have come about historically, and (iii) what it means for tone system typology, i.e. that universals are often strong tendencies but not absolute.
APRIL 28 -
Aditi Lahiri
University of Oxford / Miller Institute for Basic Research in Science, UC Berkeley
Phonological words in Germanic
There are contrary views on how phonological phrasing in Germanic is determined - either by surface syntax, by rhythmic principles or other syntax-phonology interface algorithms. After a brief survey of historical phonological developments in Germanic (having to do with cliticisation and the creation of new inflectional affixes from clitics, and with attendant changes of the forms involved), a default phonological word formation of grammatical/function words is claimed which associates them leftwards regardless of morphosyntactic constituency and which does not respect morphological word integrity either. Thus, encliticisation predominates almost exclusively, and productive inflectional affixes innovated via cliticisation are all suffixes. Experimental evidence (phonological encoding) from contemporary Germanic (in this case Dutch) confirms that phonological words are crucial units in speech planning, with such units equally formed as trochees irrespective of the phrasal syntax of grammatical and lexical words.
MAY 5 -
Qibin Ran
Nankai University
Addressing animals with clicking sounds: Patterns and universals
A strong tendency for people to produce clicking sounds to address animals is demonstrated in this paper. Based on investigations of 126 spots (67 Chinese spots and 59 non-Chinese spots) worldwide, 30 kinds of clicking sounds have been found, and all spots are observed to use clicking sounds to address at least one animal. There are 2.51 clicking sounds per spot overall. The most used clicking sounds are (top 5 and in descending order): ǀw, ǀ, ʘ, *1(retroflex rounded click), and *2(click with bilabial release and labiodental friction). Animals most frequently to be addressed with clicking sounds are (top3 and in descending order): dog, chicken and cat. Evident typological features are found meanwhile, e.g., retroflex rounded clicks (almost all are used to address chicken) and clicks with bilabial release and labiodental friction (almost all are used to address cat) are nearly only found in Chinese; while lateral clicks are only found in non-Chinese languages (especially in European languages and almost all are used to address horse). Why clicking sounds used so frequently and worldwide to address animals? We argued that the motivations are resulted from the particular characteristics and properties of clicking sounds, which sorted out as following: Clicking sounds are: (1) ideally specific to human beings; (2) are more likely to aim at animals; (3) are very prominent in perception; (4) contains adequate subtypes; (5) are enough easy to produce.
MAY 12 -
Henning Reetz
Institut für Phonetik, Universität Frankfurt
Infant cries and developmental disorders
Experienced nurses of babies are reported to be able to identify developmental problems (such as hearing impairment) and even pathological diseases (for instance, cleft palate, genetic abnormalities like Down syndrome or 22q11) from infant cries. An early diagnosis of such problems would naturally be desirable. I will present results from a project which attempts to correlate acoustic properties of infant cries with such disorders. The goal is to classify relevant acoustic properties to develop software which would determine whether a cry is linked to a particular disorder. The data is based on German babies, aged 1 day to 12 months.
MAY 21 -
Erich Round
University of Queensland
Linguistics in the age of biostatistics: how data properties matter, and the case of Bayesian clustering
Note: this talk is on a WEDNESDAY at the usual hour
Modern computational methods open up new directions of enquiry utilizing cross-linguistic datasets whose size would overwhelm traditional pencil-and-paper analyses. However, where techniques are not originally designed for linguistic research, questions will arise as to how performance might be affected by properties of linguistic datasets, whose principles of design have yet to alter much from the paper-and-pencil era. Issues such as non-independence of variables may be innocuous in some cases (Pagel & Meade 2006), but precisely when and whether this generalizes requires ongoing investigation.
In a series of experiments I examine the performance of the Bayesian clustering algorithm STRUCTURE (Pritchard et al. 2000) by making controlled variations to <10% of a large cross- linguistic dataset of typological variables (Reesink et al. 2009—121 languages; 155 binary variables; 88% of cells complete). I find that in terms of both the number of clusters inferred, and the languages they group together, STRUCTURE is indeed sensitive to the inclusion/exclusion of variables which have:
i. A high proportion of the same value across languages;
ii. A high number of missing values;
iii. A tendency to pair-wise correlation with another variable
iv. A relatively high level of correlation with all variables in the dataset
I consider two ramifications. Firstly, typological linguistic datasets do differ in some respects from the kinds of biological datasets which serve as test data for biostastical techniques. Consequently, it will be desirable to test these techniques with specifically linguistic data, to best ascertain how they perform in the linguistic domain. Secondly, a typologist’s decision over whether or not to build variables X, Y or Z into a dataset may hinge precisely on issues such as (i–iv), which can then influence the final statistical result — a manifestation of the “researcher degrees of freedom” problem (Simmons et al. 2011). Thus, to aid the evaluation of cutting-edge research, it will be desirable when datasets are designed, for the accompanying decision-making process to be documented and published, since those decisions may bear upon the eventual results, even before language coding begins.