Lisa Zsiga: Cross-linguistic Patterns in Speech Planning:
Obstruent Nasalization in Korean and Korean-accented English
As the collision of atoms provides physicists the opportunity to study basic elements that are revealed only through interaction, so the study of second-language speech production provides an opportunity to learn about the basic elements of language, by examining old patterns that persist or new patterns that are created when language systems collide. This study focuses on patterns of speech planning and production that are revealed by word-to-word timing in L2 speech. Despite the recognized importance of rhythm and phrasing in (mis-)understanding the speech of language learners, few studies have examined word-to-word coordination and external sandhi in L2 phonology. The studies that have been done have generally found that processes that apply across word boundaries do not transfer from L1 to L2, indicating that L2 speakers plan their productions one word at a time, resulting in a word “word integrity” effect (Cebrian 2000). The present study found different results.
This study examines phonetic data on one external sandhi process, Korean obstruent nasalization. In Korean, a word-final stop followed by a word-initial nasal becomes nasal: /kimpa
p mekta/ “sushi eat” becomes [kimpa
m mekta]. Do Korean speakers learning English ever pronounce English sequences according to this pattern? Word-integrity predicts that they would not.
Participants were 10 native speakers of Seoul Korean: 5 intermediate-level learners of English, 5 advanced. The participants read sentences in both Korean and English containing two-word phrases (object#verb in Korean, verb#object in English), containing targeted stop#stop and stop#nasal sequences. Examples include “stop Matt,” “take Nat,” /kimpap mekta/ “sushi eat,” and /hopak mekta/ “pumpkin eat.”
In the Korean utterances, nasalization applied in 93% of relevant tokens. For the English utterances, productions were much more varied. One participant seemed to obey word-integrity, others were able to produce English-like stop#nasal sequences. Contrary to prediction, however, nasalization often occurred: phrases like “stop Matt” were pronounced with word-final nasals (“sto[
m m]att”) 24% of the time across all the data, and as often as 70% of tokens for one participant. Perhaps most interesting is the common occurrence of voiced tokens (38%): “stop Matt” pronounced as “sto[
b m]att,” a pronunciation that is correct neither for Korean nor for English. These voiced tokens are argued to arise from transfer of the pattern of gestural overlap that gives rise to intersonorant voicing in native Korean (Jun 1996), in the absence transfer of the phonological pattern of nasalization.
Each of these different realizations is interpreted as an instance of a different pattern of speech planning, explored through the theory of coupled oscillators (Browman & Goldstein 2000, Goldstein 2008). Overall, it is argued that L2 speakers (both intermediate and advanced) do sometimes plan in units larger than the word, and may use either the phonological or phonetic plan, or both, from their L1 in pronouncing their L2.
Jonathan Bobaljik: Getting 'Better': On Comparative Suppletion and Related Topics
I present and discuss four or five universals drawn from a cross-linguistic study of comparative and superlative morphology. Special attention is given to three generalizations regarding root suppletion in the comparative degree of adjectives (good-better, bad-worse). These generalizations, I contend, have a variety consequences for morphology, semantics and perhaps syntax, particularly in the areas of lexical decomposition (at whatever level this obtains) and the formal treatment of suppletion vs. irregularity. Although comparative suppletion is rare (though attested) outside of Indo-European, and although the data sample is small within any one language, the generalizations over the total data set are surprisingly robust. Two generalizations are given here:
The Comparative-Superlative Generalization:
If the comparative degree of an adjective is built on a suppletive root/stem, then the superlative is also suppletive. The superlative may use the same root as the comparative, or may be further suppletive, but will not use the basic adjectival root. Thus the schema in (1), where A, B, C refer to phonologically unrelated roots.
(1) A - A - A completely regular: short, short-er, short-est
A - B - B suppletive: bad, worse, worst
A - B - C doubly suppletive: Latin 'good': bonus - melior – optimus
A - B - A *unattested* * bad - worse - baddest
I argue that this generalization favours analyses in which the superlative is not merely related to the comparative (e.g., both involve degree operators) but is rather _derived_from_ the comparative: [[[SHORT]-ER]-(ES)T]. Put somewhat more contentiously, I argue (with a qualification) that UG excludes a morpheme "-EST" (Superlative) that attaches directly to adjectival roots.
The Comparative-Change-of-State Generalization:
If the comparative degree of an adjective is built on a suppletive root, then a derived change-of-state verb (inchoative or causative) will also be suppletive. The verb may use the same root as the comparative (bad - worse - worsen; bonus - melior - meliorare), or may be further suppletive, but will not use the basic adjectival root.
By parity of reasoning to the first section, I must conclude (contra Dowty and others) that change-of-state verbs always include the comparative at some level of representation (cf. Kennnedy & Levin). I will defend this view against a variety of possible objections and examine apparent counter-examples.
Barbara Hall Partee: Russian Genitives, Non-Referentiality, and the Property-Type Hypothesis
Dahl (1971) observed that the same contexts often license Genitive and Subjunctive in Russian, so that the semantic contrast between Gen and Acc may be similar to that between Subjunctive and Indicative. The parallels are clearest in a paradigm due to Kagan (2005) (her paradigm, our examples):
(1) a. Ja ne zametil, čto jubilej GAI prazdnovali voditeli.
I NEG noticed that anniversary GAI celebrated drivers.NOM
‘I didn’t notice that drivers were celebrating the anniversary of the road police.’ (factive)
b. Ja ne zametil, čtoby jubilej GAI prazdnovali voditeli.
I NEG noticed that.SUBJUNC anniversary GAI celebrated drivers
‘I didn’t notice that any drivers were celebrating the anniversary of the road police.’ (non-factive)
c. Ja ne zametil vodku na stole.
I NEG noticed vodka.ACC on table
‘I didn’t notice the vodka on the table.’ (presuppositional)
d. Ja ne zametil vodki na stole.
I NEG noticed vodka.GEN on table
‘I didn’t notice any vodka on the table.’ (non-presuppositional)
The similarity between non-veridicality in the sentential domain and non-specificity in the nominal domain has been explored by Kiparsky and Kiparsky (1970), Dahl, Farkas, Giannakidou, and others. We suggest that both Subjunctive and the Russian Gen often signal the
absence of a presupposition or entailment: of truth, and of existence, respectively. Their licensing conditions are similar but not identical. A caveat (to be discussed): we do not believe that our proposed semantic generalizations apply to all cases of Object Gen Neg; the construction may be semi-syntacticized.
Kagan (2005) and Partee & Borschev (2004) propose to treat Russian alternating Genitive NPs as “property type”, the type attributed to opaque objects of intensional verbs in Zimmermann (1993) and Van Geenhoven and McNally (2005), to ‘subjects’ of existential sentences by McNally (1992), Landman (2004), and Paducheva (1985:99), to incorporated nominals in Van Geenhoven (1998), and to Russian ‘small nominals’ in Pereltsvaig (2006). In all those cases, the authors argue that such NPs lack e-type reference and bear no referential index. Arguments for and against this hypothesis for Gen Neg NPs have been given in Partee & Borschev (2007). Here we argue in favor and answer most of the arguments against. We’ll also discuss additional arguments adduced in Kagan’s 2008 dissertation.
Background readings: (1) Borschev, Vladimir, Elena V. Paducheva, Barbara H. Partee, Yakov G. Testelets, and Igor Yanovich. 2008. Russian genitives, non-referentiality, and the property-type hypothesis. In Formal Approaches to Slavic Linguistics: The Stony Brook Meeting 2007 (FASL 16), eds. Andrei Antonenko et al. Ann Arbor: Michigan Slavic Publishers.
https://udrive.oit.umass.edu/partee/FASL16FinalForPrinting.pdf
(2) Partee, Barbara H., and Borschev, Vladimir. 2004. The semantics of Russian Genitive of Negation: The nature and role of Perspectival Structure. In Proceedings from SALT XIV, ed. Robert B. Young, 212-234. Ithaca: CLC Publications.
http://people.umass.edu/partee/docs/SALT14ParteeBorschev.pdf
--------------------------------------------------------------------------------------
*
Note: This talk is based mainly on Borschev et al (2008), reading (1) above. Longer versions in English and in Russian are in progress; comments will be most welcome.
Lyle Campbell: How Many Language Families are there in the World? How Many Language
Isolates?
How many language families are there in the world? How many language isolates? Most historical linguists do not know, and opinions about the number of language families vary from only one to around 400, ranging around 250. Linguists agree only that there is no agreement on the number. The answer to this question is complicated by several theoretical and methodological questions currently debated:
(1) Is the family tree model flawed?
(2) Can linguistic diffusion be a serious challenge to linguistic genealogical relationships?, or, put differently, how successful in difficult cases can we be at distinguishing inheritance from borrowing?
(3) Is there a temporal threshold beyond which genetic relationship among languages is no longer demonstrable?
(4) Why have typological comparisons sometimes led to erroneous hypotheses of language families?
(5) Why is it that human genetics and other non-linguistic fields cannot help us determine genetic relationships among languages?
(6) Can the speakers’ type of society or their sort of culture determine the nature of change in their languages and thus influence how they might be classified?
(7) What is the prognosis for discovering new family relationships among languages?; what recent progress has been made?
This talk considers the topic broadly and addresses these questions, attempting to proffer answers to (or at least perspective on) some of them and thus to contribute to language classification generally, and to contribute ultimately to answering the questions, how many language families are there, really, and how many language isolates are there?
William Schuler: A Simple Computational Model of Interactive Language Processing
Psycholinguistic studies (e.g. by Tanenhaus and colleagues) suggest a model of human language processing in which parsing and word recognition interactively depend on the interpretations of hypothesized words in some discourse or environment context. At the same time, it is attractive to assume that many kinds of intra-sentential processing (including parsing and semantic composition) are performed within a severely-constrained short-term memory store, possibly containing as few as three or four disconnected elements. This talk will describe an implemented computational model of language processing that attempts to satisfy both these desiderata, tempering the rich conditioning of the interactive model with austerity of human-like memory bounds. The result is a simple, implementable, and empirically-successful model that provides a unified explanation of some previously unrelated linguistic phenomena.
The model parses sentences in a transformed representation that minimizes memory requirements by turning right-branching structures into left-branching structures of incomplete constituents, which can be composed immediately upon being encountered. The talk will first present results of a corpus study suggesting that this transformed constituent structure is sufficient to allow a vast majority of naturally occurring sentences to be parsed using a memory store containing only three or four incomplete constituents. I will then present evidence that this constituent structure also provides a natural explanation of speech repair (in which speakers `back up and restart' incomplete constituents following speech errors) as an ordinary syntactic phenomenon, providing accuracy comparable to that of the best unbounded-memory parsers. Finally, I will describe a simple factorization of this model to allow interactive interpretation, by augmenting each incomplete constituent in the memory store with an `incomplete referent': a vector of word co-occurrences, latent concepts, or sets of individuals, composed through matrix multiplication. Results using a real-time speech interface implementing this interactive model show this model can deliver more accurate recognition than a non-interactive baseline.
Two articles (the former under review, the latter in press) describing the syntax and semantics of this model are available on my web site:
http://www-users.cs.umn.edu/~schuler/paper-jcl08wsj.pdf
http://www-users.cs.umn.edu/~schuler/paper-jcl07slush.pdf
Seth Cable: Towards the Elimination of Pied-Piping: Evidence from Tlingit
This talk presents evidence that pied-piping, as traditionally
understood, might not exist. That is, I propose that movement only
ever targets the maximal projection of the lexical item whose features
trigger it. Classic, well-known counter-examples from English are
reanalyzed in light of new data from Tlingit, an understudied and
endangered language of Alaska. It is argued that the prima facie
appearance of ‘pied-piping’ structures in Tlingit is misleading, and
actually reflects structures where no true pied-piping has taken
place. I then turn my attention to putative cases of pied-piping in
English, and argue that a similar ‘eliminativist’ analysis is possible
here. The overall picture that results is one where both the
phenomenon of ‘pied-piping’ and the grammatical mechanisms introduced
to derive it are eliminated from the theory of grammar.
Asaf Bachrach: fMRI investigation of incremental language processing in a
naturalistic context
Our study examined brain activation in a naturalistic language
processing task, with a particular focus on the temporal dynamics
inherent to this complex cognitive task.
Sentence processing, in particular in the auditory modality, is
incremental. The structure and associated compositional meaning of a
sentence are not provided to the listener instantaneously, but require
integration over multiple temporally spaced inputs. Behavioral and
electrophysiological evidence (a small sample of which will be reviewed)
point out that the human parser makes use of an `eager' strategy,
incrementally constructing the eventual sentential representation based
on partial input. In addition, it appears that this incremental strategy
is probabilistic and parallel. The parser considers potentially multiple
alternative analyses, which are probabilistically weighted.
Most behavioral and imaging paradigms used to explore aspects of
incremental auditory sentence processing have been limited by the use of
qualitative or binary contrasts and by a sparse sampling approach (often
only one data point per sentence). In this talk we will present the
results of a novel imaging paradigm that attempts to overcome the above
limitations.
We used functional Magnetic Resonance Imaging (fMRI) to monitor brain
activation while subjects passively listen to short narratives. The
texts were written so as to introduce various syntactic complexities
(relative clauses, embedded questions, etc.) not usually found (in such
density) in actual corpora.
With the use of computationally implemented probabilistic parser (taken
to represent an 'ideal listener') we have calculated a number of
temporally dense (one per word) parametric measures reflecting
different aspects of the incremental processing of each sentence. We
used the resulting measures to model the observed brain activity (BOLD).
We were able to identify different brain networks that support
incremental linguistic processing and characterize their particular
function.
In the talk we will present data regarding the effect of contextually
based prediction (or surprisal), distinguishing lexical and syntactic
prediction, and the effect of local structural ambiguity, distinguishing
the effects of uncertainty from these of reanalysis.
Edward Stabler: Greedy Incremental Parsing
Fluent speech is typically understood almost word-by-word, with
syntactic and other ambiguities resolved almost immediately (as
evidenced by Tanenhaus et al, for example). This is often taken as
evidence against modular theories in which syntactic structure is
assembled based on syntactic properties. But this raises worries about
what a non-modular syntax could look like, and often encourages the
view that generative grammar is at best only very abstractly related
to the structures recognized in fluent speech. This is sometimes taken
to be confirmed by the variety of proposals in generative grammar, and
by acquisition theories according to which generative grammars fail to
characterize what people learn (as in Tomasello, for example.)
This talk will show that these reactions are inappropriate,
with particular attention to approaches to incremental recognition of
relative clause structure: non-modular approaches represent a step
away from theories that can account for the facts, not a step towards
them. In particular, (1) although there is great diversity in detail
among contemporary generative grammars, a very substantial consensus
on fundamental properties has emerged within and across grammar
traditions, a consensus that is gradually becoming clearer and more
explicit (fundamentals of 'UG'); (2) structures defined by these
consensus grammars can be recognized with simple and well-understood,
efficient methods; (3) standard recognition methods allow incremental
assessment of the current parse at a very fine grain, so non-syntactic
factors can select among the (modular, syntactically defined) options
"immediately"; (4) Some psycholinguistic proposals suggest a kind of
"greedy" parsing strategy which may be rather close to the standard,
well-understood methods.
John Hale: Dependency Grammar in a Computational Model of Human Sentence Parsing
How do human readers comprehend sentences? This talk presents a cognitive model of sentence comprehension based on word-to-word connections. Relating the concepts "lengthy cognitive process" and "surprising automaton transition,” the derived predictions help explain eye-fixation durations in ways that linguistically impoverished models cannot. This highlights the role of sentence structure in comprehension and points to the slippage between ideal and human-like parsing. [Joint work with Marisa F. Boston (Cornell University), Reinhold Kliegl, Umesh Patil, and Shravan Vasishth (Potsdam University).]
Jens Michaelis: Locality conditions from a mildly context-sensitive perspective
In the wake of Peters and Richie's demonstration in 1973 that transformational grammar is Turing-equivalent, there was an intensive search for ways to reduce expressive power using locality conditions (LCs). Complexity results, however, remained largely absent. With Stabler's formalization of minimalism (Stabler 1997), this picture has changed: Minimalist Grammars (MGs) are mildly context-sensitive in the sense of Joshi (1985), in part, because they impose an LC called the shortest move condition.
The talk explains how the complexity of MGs is affected by the presence or absence of various LCs, demonstrating that LCs as such are not automatically restrictive where a formal notion of restrictiveness is applied.
Paul Hagstrom: Intervention effects and the flavors of Q
In Japanese, questions are formed with the help of a question particle
"ka" ("Q"), a morpheme that can also be used to signal disjunction or
combine with a wh-word like "dare" 'who' to form "dareka" 'someone'.
This pattern (with some variations) recurs across many languages.
Familiar comparisons with Sinhala and earlier stages of Japanese
itself lead to the hypothesis that this "Q" particle undergoes
syntactic movement in questions (a proposal much like that recently
presented by Seth Cable). We can hypothesize a coherent semantic
contribution for each of these pieces, which leads to the more general
hypothesis that this process is universal across languages (being part
of the means of forming a question or indefinite meaning). One
prominent argument for the syntactic movement comes from "intervention
effects" (that prohibit certain "intervenors" from the path between a
wh-phrase in situ and its associated interrogative complementizer).
In this talk, I will tackle two aspects of this project where open
questions remain. The first concerns the identification of Q
morphemes across languages: many languages seem to make finer
morphological distinctions than Japanese does, suggesting more than
one "flavor" of Q. The second concerns the analysis of the
intervention effect itself, in response to several recent proposals
attributing them to semantic, pragmatic, or prosodic causes, rather
than to syntactic causes.
Cecilia Ovesdotter Alm: Affect in Text and Speech
As technology and human-computer interaction advance, there is an
increased interest in affective computing and socioemotive aspects of
language. One of the current challenges in computational speech and text
processing is addressing affective and expressive meaning, an area that
has received fairly sparse attention in linguistics. Linguistic
investigation in this area is motivated both by the need for scientific
study of subjective language phenomena, and by useful applications such as
expressive text-to-speech synthesis.
The presented work makes several contributions to the study of affect and
language. After describing a new sentence-annotated data resource for
large-scale exploration of affect, the talk discusses characteristics of
a high agreement data subset authored by H. C. Andersen. It also outlines
models and challenges for exploring affect in language and reports on
empirical results from automatic affect prediction in text as well as the
use of interactive evolutionary computation to evolve emotional prosody.
The talk concludes with identifying and suggesting paths for further
exploration.
Patricia Keating: Prosodic position and segmental articulation in English
Previous research has shown a close link between segments' phonetic
properties, and their position in prosodic structure. This link has been
described as a strength relation: segments in stronger prosodic
positions have stronger phonetic realizations. Prosody serves both a
grouping function and a prominence-marking function in speech. Some
prosodic positions are strong because they are initial in a prosodic
domain; others are strong because of prominence. Are strengthenings
associated with initial positions and with prominences the same? In a
study with Taehong Cho (now at Hanyang University), we compared the
effects of utterance-initial position and two levels of prominence in
English, lexical primary stress and phrasal accent. We measured several
articulatory (electropalatographic) and acoustic properties of
word-initial CV syllables. In our data, boundary effects and prominence
effects were seen in different phonetic properties, or on the same
properties but with effects in opposite directions. Boundary effects
were seen primarily in consonantal measures, while most effects on
vowels were due to prominence. However, domain-initial strengthening
seems to be constrained by prominence: in many cases the effect of
domain-initial strengthening was greater when there was no concomitant
effect of stress or accent. Finally, the temporal scope of phrasal
accent on a word could be examined in these data: is it the stressed
syllable or the whole word that manifests an accent? While the effects
of accent were strongest on the primary-stressed syllable of an accented
word, several phonetic measures revealed accent effects on
secondary-stressed syllables. We conclude that domain-initial
strengthening is not the same thing as prominence-related strengthening.
Timothy O'Donnell: Language Reuse and Computation
Productivity in language is made possible by a division of labor between computation and storage: stored lexical items are combined via computation into more complex structures. A central question for theories of language is what constitutes this inventory of stored items: Where do the stored items come from? Under what conditions does storage happen? How are storage and computation integrated? I will present a Bayesian framework designed to study these questions, along with some preliminary empirical evaluation.
John McCarthy: Harmonic Serialism
This talk examines a version of Optimality Theory that incorporates serial derivations. Called Harmonic Serialism (HS), this approach was introduced, briefly discussed, and ultimately set aside by Prince and Smolensky (1993/2004).
The candidate-generating component Gen is the main locus of difference between “classic” OT and HS. In classic OT, Gen can produce candidates that differ from the input in multiple ways: e.g., Gen(/pat/) = {pat, pati, pad, padi, peti, …}. HS’s candidates are allowed only one difference from the input at a time: Gen(/pat/) = {pat, pati, pad, …}. In HS, the Eval component selects the optimal member of this limited candidate set, which then becomes the input to another pass through Gen and Eval. This process continues until convergence: the input to Gen and the output of Eval are identical. In HS, the path to ultimate optimality is gradual rather than immediate.
When the input and ultimate output differ in two or more ways, HS requires two or more passes through Gen and Eval, whereas classic OT always requires just one. To get from /pat/ to [padi], classic OT requires a constraint ranking in which [padi] is more harmonic than [pat], [pad], [pati], etc. HS has another requirement: some form intermediate between /pat/ and [padi], either [pati] or [pad], has to be more harmonic than [pat] and less harmonic than [padi]. If plausible constraints and possible rankings don’t provide such a form, then the mapping from /pat/ to [padi] will be impossible in HS but possible in classic OT,
ceteris paribus.
Is this property of HS supported by the facts? To answer that question, I’ll discuss some examples of the “too-many-solutions” (also known as “too-many-repairs”) problem. This problem arises whenever ranking permutation predicts unattested ways of satisfying a markedness constraint.
