Cornell University Linguistics Department

Janet Pierrehumbert: Sustaining Linguistic Complexity

The central role of frequency in the acquisition and use of linguistic patterns is abundantly supported by a wide range of linguistic and psycholinguistic results. However, a simple baseline model of frequency effects has poor limiting behavior. In the simple baseline model, learners acquire the frequencies of patterns from the rate at which they encounter these patterns while interacting with mature members ofthe linguistic community. Their productions mirror the frequencies of exposure. For aspects of the language system involving finite numbers of distinct elements (such as phoneme inventories or verb paradigms), the language system is predicted to become simpler and simpler over time due to the combination of random walking and undersampling effects.

This talk will analyze cases in which complexity persists for an unexpectedly long time. What mechanism supports the continued existence of rare patterns? I will begin with a simple study of "Zipf's Law" for phoneme inventories. The main example will be a study of arbitrary morphological gaps in Russian verb conjugation, analyzed using a Bayesian population-level learning model.

Kemmonye Monaka: An acoustic study of Shekgalagari stops

Researchers have been divided on whether the glottalic or pulmonic airstream mechanism is used in the production of the voiceless unaspirated stops in Shekgalagari and related languages. Decisions have hitherto been based only on auditory impressionistic methods. This study aims at investigating the phonetic realisation of the voiceless unaspirated plosives in Shekgalagari from recordings of 4 native speakers who served as linguistic consultants. The investigation is done by means of four synchronised channels of data: Gx, Lx, Sp and the spectrogram. Mostly qualitative, but also quantitative information on the stops was derived. The Gx, which depicts the displacement of the larynx in the egressive and ingressive airflow mechanism, displays a relatively level signal, indicating an undisplaced larynx for airflow initiation. The Lx signal and the spectrogram depict modal phonation prior to and after the plosives, indicating lack of glottalization. The burst in the speech signal is weak and hardly visible in some cases, indicating lack of significant air pressure in the vocal tract. VOT values are small and vary as a function of articulatory place. These results show that the stops appear to be plain voiceless unaspirated stops rather than ejectives as they have sometimes been described. Further investigation on closure and total duration, amplitude of the burst, vowel pitch and vowel rise time need to be done to help shed light on the nature of production of the voiceless unaspirated plosives in Shekgalagari.

Heidi Harley: Applicatives, causatives, and the Mirror Principle: vP-external subjects and Voice

The interaction of applicative, causative, and passive morphology on verbs in Hiaki (Yaqui), a Uto-Aztecan language of northern Mexico and southern Arizona, reveals that VoiceP and vP must be independent projections, given a syntacticocentric view of morphological composition. In this talk, I first present the constellation of evidence that prove the necessary independence of VoiceP and vP in Hiaki, and then discuss the implications for the analysis of Voice phenomena cross-linguistically, including noncausative reflexive morphology in Romance.

Meghan Clayards: Acoustic-phonetic speech cues and word recognition: what makes a good cue?

The speech signal is made up of multiple, probabilistic acoustic-phonetic cues. A central question in understanding speech perception and word recognition has been how those probabilistic cues map onto mental representations of phonemes or words. One aspect of the question is how to deal with a variable or probabilistic cue. Is the variability noise or is it informative? The second aspect is how to combine multiple cues. Should each cue be treated equally? Are some cues primary and others secondary? I propose an account which addresses all of these issues by using the distributions of the cues themselves. My account formalizes the intuition of phoneticians that the separability of speech categories by acoustic-phonetic cues should determine the usefulness of the cue for perception. I will present both production and perception data relating to how acoustic-phonetic cues are used by speakers and listeners.

Florian Jaeger: Optimal Language Production: Uniform Information Density

Theories of rational cognition (Anderson 1990; Simon 1990; also Bayesian approaches to cognition) consider the brain to be close to optimal in terms of efficient information processing. In language research, the idea of optimal information processing has received relatively little attention (but see recent work by Hale 2001; Genzel & Charniak 2002; Aylett & Turk 2004; Levy 2006; Jaeger 2006). I present several studies that investigate the extent to which language production is optimal.

Speakers often have a choice among several ways to convey the same message, e.g. in cases of optional reduction:

(1)	Morphosyntactic reduction
	I'm ...	vs.	I am ...
	He wouldn't ...	vs.	He would not ...
(2)	Syntactic reduction
	He thinks I am a loser.	vs.	He believes that I am a loser.
	The guy known to be ...	vs.	The guy that is known to be ...

Information theoretic considerations lead to the prediction that optimal speakers should prefer choices that keep the amount of information¹ conveyed per time uniform (the hypothesis of Uniform Information Density, Jaeger, 2006; Levy & Jaeger, 2006). For cases of reduction, as in (1) and (2) above, Uniform Information Density predicts that speakers are more likely to choose the full form (e.g. am) rather than the reduced form (e.g. 'm) the more information the reducible word contains.

I present corpus studies supporting the hypothesis of Uniform Information Density for several cases of syntactic and morpho-syntactic reduction in spontaneous speech. Modern statistical methods (mixed logit models) are used to analyze the highly unbalanced and clustered data typical for corpus studies, while controlling for other factors known to affect reduction. I also briefly discuss further evidence from the word order choices in production, the distribution of disfluencies, and phonetic reduction. I conclude that Uniform Information Density captures an important generalization of the computational system underlying language production and that this system is - at least to some extent - optimized.

¹Information theory defines the Shannon information content of a unit (e.g. a word) in terms of its probability. Specifically, the information of a word corresponds to the logarithm of the inverse of the probability of the word in its context. This makes it possible to estimate the information of a word from corpora.

B. Elan Dresher: Strange but Indubitable: Against Phonetic Determinism of Contrast in Phonology

Phonological contrast deals with contrastive and redundant features. It is not the same as phonetic contrast, which deals with perceptual distinctiveness. I will first present a theory of how to assign contrastive representations, then I will give examples of how this theory can contribute to accounts of phonological patterning.

Phonological contrasts are not self-evident. Phonologists have employed two distinct and inconsistent methods for assigning contrastive features. One common but flawed method works by making pairwise comparisons of fully-specified phonemes, throwing out logically redundant specifications. I have argued that contrast is determined by a hierarchy of features. Features are ordered and assigned to phonemes successively until all phonemes have been differentiated; features not assigned by this process are redundant. Different orders result in different contrastive specifications.

Phonological contrast as assigned above plays a central role in understanding phonological systems. Trubetzkoy (1939) presents many examples where the patterning of a phoneme depends not on its phonetics alone, but on its place in the system of contrasts, which is itself determined, it is argued here, by the feature hierarchy of that language. Similarly, Sapir (1925) considered it 'most important to emphasize the fact, strange but indubitable, that a pattern alignment does not need to correspond exactly to the more obvious phonetic one.' In the second half of the talk I will present a series of case studies, drawing on the work of members of our project in Toronto done over the past fifteen years (http://www.chass.utoronto.ca/~contrast). These studies cover different types of phonological processes, including vowel harmony in Manchu-Tungusic, Mongolian, and Turkic (Zhang 1996), consonant co-occurrence in Bumo Izon (Mackenzie 2005), and adaptation of English loan words into Polynesian languages (Herd 2005). They all show the importance of phonological contrast, as instantiated by the contrastive hierarchy. At the same time, they show the insufficiency of purely phonetic approaches to phonological phenomena.

References

Herd, Jonathon (2005). Loanword adaptation and the evaluation of similarity. Toronto Working Papers in Linguistics (Special Issue on Similarity in Phonology) 24: 65-116.
Mackenzie, Sara (2005). Similarity and contrast in consonant harmony systems. Toronto Working Papers in Linguistics (Special Issue on Similarity in Phonology) 24: 169-182.
Sapir, Edward (1925). Sound patterns in language. Language 1: 37-51. Reprinted in Martin Joos (ed.), Readings in Linguistics I. Chicago: University of Chicago Press, 1957, 19-25.
Trubetzkoy, N. S. (1939). Grundzüge der Phonologie. Göttingen: Vandenhoek & Ruprecht.
Zhang Xi (1996). Vowel Systems of the Manchu-Tungus Languages of China. Doctoral dissertation, University of Toronto.

Marc Brunelle: Vietnamese tone: featureless phonology and phonetic coherence

Vietnamese tones can be grouped into classes that are active in phonological alternations. However, repeated attempts to define these classes in phonetic terms have been unsuccessful (Earle 1975, Gsell 1980, Burton 1982, Ng\364 1984, Alves 1995, Pham 2001, 2003, Idsardi and Avery ms). A perceptual study was therefore conducted to uncover the acoustic properties to which native listeners are sensitive. This study confirms that the phonological classes cannot be associated to phonetic properties, but unexpectedly reveals that the perception of Vietnamese tone is nontheless structured and organized along a small and economical set of acoustic cues. Potential implications for feature and gesture economy are discussed (Clements 2003).

John Hale: Quantifying Sentence Processing with Information Theory

The relationship between grammar and language behavior is not entirely clear-cut. For instance, in comprehending sentences, people manifestly do computational work. But what is the relationship between difficulty on a particular sentences and an individual's more general knowledge of grammar?

This talk presents one such relationship, the Entropy Reduction Hypothesis (ERH). Combining grammar and information theory, the ERH formalizes the intuition that people have more difficulty on words that do more work for the comprehender.

The general theory is tested using explicit grammar fragments including Minimalist Grammars (Stabler 1997) expressing rival transformational analyses of relative clauses.

The theory will be shown to derive a range of well-documented processing phenomena including garden-path sentences, center-embedding, and the Accessibility (or Obliqueness) Hierarchy.

Edward P. Stabler, Jr. 1997. ``Derivational Minimalism'' in Logical Aspects of Computational Linguistics, edited by Christian Retor\351. Springer-Verlag.

Roger Levy: Probabilistic knowledge in human language comprehension and production

The talk covers two fundamental issues, one each in language comprehension and production: what determines the difficulty of comprehending a given word in a given sentence, and what factors influence the choice that a speaker makes when it is possible to express a meaning more than one way? The first half of the talk presents a theory of processing difficulty, building on Hale (2001), derived from the premise that sentence comprehension involves the rational and fully incremental application of probabilistic knowledge. On this theory, the comprehender's probabilistic grammatical knowledge determines expectations about the continuations of a sentence at multiple structural levels; and these expectations determine the difficulty of processing the words that are actually encountered. I show how this theory can be applied to a number of results in online sentence comprehension (Konieczny, 2000; Konieczny & Doering, 2003; Jaeger et al., 2005) that are problematic for memory-oriented theories such as the Dependency Locality Theory (Gibson 1998, 2000) or Similarity-Based Interference (Gordon et al., 2001, 2004; Lewis et al., 2006), yet were not covered by previous probabilistic theories because the results do not involve resolution of structural ambiguity.

The idea that probabilistic expectations drive processing difficulty leads to the proposal of the second half of the talk: that speakers make choices in language production such that their utterances tend toward an optimal, uniform level of information density. The second half of the talk introduces the basic theory of uniform information density, and presents an empirical study and model using the parsed Switchboard corpus to investigate speaker choice in optional relativizer omission, such as (1) below:

How big is the family (that) you cook for __?

We find that speakers tend to use the optional relativizer "that" more often when the information density of the onset of the relative clause is higher. These results provide evidence in support of uniform information density as a locus of optimal production decisions.

Alexander Koller: Natural Language Generation as Planning

The problem of natural language generation is intimately related to AI planning on many levels. In both problems, the computer has to search for a sequence of actions that combine in appropriate ways to achieve a given goal; in the case of generation, these actions may correspond to uttering speech acts, sentences, or individual words. This has been recognized in the literature for several decades, but there is currently a revival of interest in exploring this connection, which has been sparked especially by the recent efficiency improvements in planning.

In my talk, I will first show how sentence generation can be translated into a planning problem. This has the advantage that the (somewhat artificial) separation of sentence generation into microplanning and surface realization can be overcome. Furthermore, each plan action captures the complete grammatical, semantic, and pragmatic preconditions and effects of uttering a single word. I will then present a recent proposal for a shared task for the generation community, in which the system must generate instructions in a virtual environment. I will discuss some problems that arise in this application -- particularly regarding the use of extralinguistic context in generation -- and propose some ideas on how they can be tackled using a planning approach.

Sharon Goldwater: From Sounds to Words: Bayesian Modeling in Early Language Acquisition

The child learning language is faced with a difficult problem: given a set of specific linguistic observations, the learner must infer some abstract representation (a grammar) that generalizes correctly to novel observations and productions. In this talk, I argue that Bayesian computational models provide a principled way to examine the kinds of representations, biases, and sources of information that lead to successful learning. As an example, I discuss my work on modeling word segmentation. I first present a computational study exploring the effects of context on statistical word segmentation. In this study, a model that assumes words are statistically independent (as in the stimuli used in many human experiments) is compared to a model that defines words as units that help to predict following words. I show that the context-independent model undersegments the data, while the contextual model yields much more accurate segmentations, outperforming previous models on realistic corpus data. This difference suggests the need to consider contextual effects in infant word segmentation.

Simulations using corpus data provide insight into the kinds of information that are useful for learning, but it is also important to address the question of whether model predictions are consistent with human learning patterns. In the second part of this talk, I present results from a project designed to evaluate the predictions of various word segmentation models. The human data is based on experiments similar to those of Saffran et al. (1996), but several parameters of the stimuli were varied between subjects to modify the difficulty of the task. The Bayesian model described above correlates better with human patterns of difficulty than any other model tested, suggesting that this model does indeed capture important properties of human segmentation.

Dorit Abusch: Temporal and Circumstantial Dependence in Counterfactual Modals

This paper examines the hypothesis, which has been advanced by Mondadori and Condoravdi, that the modality in counterfactual uses of "might" and "could" in sentences such as "he might have won, but unfortunately he didn't" is a metaphysical modality which is characterized in terms of branching time. I present counterexamples to the branching-time analysis, together with more conceptual objections. Further I argue that the truth of such sentences is context-dependent in a way which the branching-time analysis does not explain. My own analysis involves a circumstantial modality which is formalized using a time-dependent premise function.

Nabila Louriz: The phonetics-phonology interface in Moroccan Arabic Loanword Adaptation

Whether loanword adaptation is phonologically or phonetically motivated is a highly debated issue. This paper is an attempt to investigate how phonology and phonetics can account for adaptation patterns employed to accommodate French vowels in loanwords in Moroccan Arabic (MA, hereafter). The focus is on (i) front unrounded vowels and (ii) nasal vowels, which are absent from the vowel inventory of MA.

As far as the adaptation of nasal vowels is concerned, previous research claims that –in languages that lack phonemic nasal vowels- they are repaired as a sequence of oral vowel + nasal consonant (VN, henceforth). That is the nasal vowel of the donor language undergoes the process of “unpacking” and is adapted as VN in the host language (Paradis & Lacharite 1996, Paradis and Prunet 2000, Rose 1999, to name but a few). This follows from languages’ tendency to preserve borrowed phonological information. However, data from MA show that the nasal vowel is repaired in three ways: (i) VN (ii) V, or (iii) deleted altogether. This is demonstrated in the following examples.

MA	French	Gloss
gufel	Gonfler	Swell
klakson	Klakson	Horn
kwansa	Coincer	To block
zbiktur	Inspecteur	Inspector
kofra	Coup-franc	Out-of-Boards
fran	Frein	Brake
(1)

Examples in (1) display irregularities in nasal vowel adaptation, which formal phonological theories are clearly unable to account for.

On the other hand, previous research shows that the vowel /ü/ is repaired either as [i] or/and [u] in languages that lack front unrounded vowels. This repair is evidence for the monophonemic nature of /ü/ (Paradis & Prunet, 2000). However, no explanation is provided to why it is repaired as [i] and [u] in such languages as MA, where the French front unrounded vowels /ü/ displays irregularities in adaptation. For the sake of illustration, consider the following examples:

MA	French	Gloss
kustim	Costume	Suit
minut	minute	Minute
figura	Figure	Figure
furir	fourrure	Fur
bifi	buffet	Buffet
(2)

Here, /ü/ is adapted either as [i] or [u]. The observation is that the repair as [i] takes place in the environment of a front unrounded vowel, and [u] in the vicinity of a back rounded vowel. Again, this cannot be readily accounted for in phonological terms per se.

I shall explain why phonological explanation per se is insufficient. Then I shall discuss how both phonology and phonetics are brought into play to repair non-existent vowels in the host language. I shall conclude by presenting an analysis that can account for the different strategies manifested in the adaptations above, namely, one that incorporates both phonology and phonetics.

Augustine Agwuele: Signal variability: Locus equation metrics

A persistent and critical issue in accounting for the production and perception of CV sequences is how to deal with the context-induced variability that is part and parcel of the speech signal. Locus equations consistently illustrate the lawfulness of F2 transitions at the categorical level (Sussman et al 1991). The success of locus equations (LEs) lies in their ability to provide a relational invariance at a categorical level; thereby resolving vowel context-induced variability of F2 transitions so evident when individual tokens are compared.

In exploring the functional relevance of the lawfulness of locus equations, Sussman (2002) has used the same principles neuroethologists evoke to explain the functioning of neural columns coding highly variable visual (macaque) and auditory (barn owl) information. "Hypothetical auditory columns for encoding the vowel context induced variability of the F2 transitions are described using acoustic data from locus equation scatterplots of human speech. Such analyses serve to absorb the allophonic variability inherent in coarticulated speech utterances" (ibid). The variability, thus absorbed by the LEs, is probably 'masked' by the self-organization of scatterplots of stop place categorical equivalence classes. The LE does not seek to capture a consonant onset independent of the vowel context, rather it highlights that F2-CV-onset is a function of F2 values at the midpoint, and this relationship systematically varies depending on stop place. What happened to the allophonic variations that are absorbed by the locus equations? This talk will examine the responsiveness of locus equations to speech perturbations along the hyper and hypo continuum and provide a methodology to pry apart and quantify the variability 'absorbed' by Les.

William Idsardi: Phonological Poverty

Chomsky's argument from the poverty of the stimulus (POTS, Chomsky 1980) has attracted a lot of attention (pro and con) in syntax, but not in phonology. The phonologists who have mentioned it (e.g. Blevins 2005, Carr 2006) claim that it is impossible to have POTS arguments in phonology. Following Chomsky and Fodor (1980) we first explore the relation between POTS and other problems of induction, including recent results from statistical learning theory. We then present a thought experiment involving two kinds of phonological processes: stress and devoicing, showing how the environments for such processes differ typologically in ways that do not seem easily discoverable from the data nor from general considerations of simplicity. We then discuss results from artificial language learning experiments confirming these typological derived predictions. We conclude that there are POTS arguments for phonology, and that different types of phonological processes require different innately-specified learning mechanisms.