John Hale: What a Rational Parser Would Do
If we conceptualize a theory of human sentence comprehension as a combination of
(1) a grammar (2) a strategy for using the rules of the grammar and (3) some architectural facilities like memory
we still have a huge space of possible theories. It would be nice to narrow this class down to just those that
somehow made sense in relation to the communicative function sentence-comprehension often serves.
This talk examines a smaller class of comprehension theories that strive to finish parsing as soon as possible.
These theories would be ``rational'' on a view of the comprehender as doing his or her best to understand what the speaker means.
I shall argue that they correctly derive well-known garden pathing phenomena along with the puzzling Local Coherence
effects studied by Tabor, Galantuccia and Richardson (2004). Time permitting, I will discuss the relationship between this
class of theories and the Entropy Reduction Hypothesis revived in Hale (2006).
Tabor, W., Galantuccia, B., & Richardson, D. (2004). Effects of merely local syntactic coherence
on sentence processing. Journal of Memory and Language , 50 (4), 355-370.
Hale, J. (2006). Uncertainty about the rest of the sentence. Cognitive Science , 30 (4), 609-642.
Mark Baker: Two Modalities of Case Assignment: Case in Sakha
Two distinct ideas about how morphological case is assigned exist in the recent generative literature: the standard Chomskian view that case is assigned by designated functional heads to the closest NP via an agreement relationship, and an alternative view in which case is assigned to one NP if there is a second NP in the same local domain (Marantz 1991). I claim that these two ways of assigning case are complementary, based on data from the Turkic language Sakha. Accusative case and dative case in this language are assigned by Marantz-style configurational rules that do not refer directly to functional categories. This is shown by evidence from passives, agentive nominalizations, subject raising, possessor raising, and case assignment in PPs. In contrast, there is evidence that nominative and genitive are assigned by functional heads in the Chomskian way, as shown by the distribution of nominative case and the relationship between case marking and agreement. The two methods of case assignment thus coexist, not only in Universal Grammar, but even in the grammar of a single language. This raises the interesting question of how these two modes of case assignment are distributed typologically in languages of the world.
Tim O'Donnell: Bayesian models of language and stochastic functional programming
Traditionally, linguists have perceived a dichotomy between highly structured rule-based models and "statistical" approaches to language (Pereira 2000). However, over the years computational linguists and others have developed probabilistic models of language built on top of rich representations. More recently this approach has spread to other areas of cognitive science and psychology in the form of so-called "structured Bayesian modeling." Structured Bayes is one approach for allowing probabilistic models to be defined over traditional logical systems such as first-order logic or the lambda calculus. This Fall, Cornell Linguistics hosts Timothy O'Donnell (Harvard) and Noah Goodman (MIT) as they present a pair of tutorial workshops designed to introduce structured Bayesian modelling and its applications to language. O'Donnell and Goodman will present methods for defining and reasoning with Bayesian hierarchical generative models. They will do so using "Church", a programming language designed for succinctly and elegantly expressing Bayesian models.
A wiki for the Church project can be found
here.
Funded by a Small Grant from the Cornell Institute for the Social Sciences.
Fernando Pereira. Formal Grammar and Information Theory: Together Again?
Philosophical Transactions of the Royal Society 358 1239-1253 (2000)
Fragment Grammars: exploring Computation and Reuse in Language.
Timothy J. O'Donnell, Noah D. Goodman, and Joshua B. Tenenbaum
Technical Report MIT-CSAIL-TR-2009-013
http://hdl.handle.net/1721.1/44963
Church: a language for generative models. N. D. Goodman, V. K. Mansighka, D. Roy, K. Bonawitz, J. B. Tenenbaum (2008). Uncertainty in Artificial Intelligence 2008.
Beth Levin: When in Means 'into': Implications for a Typology of Motion Events
Talmy's influential proposal that languages fall into two types with
respect to the lexicalization of motion events---path languages and
manner languages---is now considered too simplistic: many languages,
though predominantly of one type, may show properties of the other.
Among these are the path languages French, Italian, and Spanish, which
allow a handful of manner of motion verbs to take PP complements in
the expression of directed motion, counter to their Talmyan type.
Although this exceptional behavior has been accommodated through
specially annotated lexical entries for the relevant verbs, I propose
a pragmatic account is preferable on empirical and conceptual grounds.
Several recent studies trace differences among path and manner
languages to their preposition inventories (Cummins 1998, Jones 1983,
Song 1997, Son 2007). English is more permissive in the expression of
directed motion because
to allows goals to be semantically composed
with manner of motion verbs. In contrast, Romance languages lack a
dedicated goal preposition;
a, though often glossed `to' in
translations of motion event descriptions, is inherently locative and
best glossed `at'. Thus, this preposition alone is typically unable
to predicate a result location of manner of motion verbs---the
hallmark of a manner language.
The problematic Romance examples, then, are better described as having
locative PPs with a directional interpretation. Interestingly,
precisely this phenomenon is observed in English and some other
Germanic manner languages (Gehrke 2008, Nikitina 2008, Thomas 2004,
Tungseth 2008), as in
Jill quickly walked in the kitchen. Nikitina
(2008) argues that such interpretations arise when adequate pragmatic
contextual support is available. Drawing on corpus studies (Baicchi
2005, Kopecka 2009, Martinez Vazquez 2001), I extend this pragmatic
explanation to Romance.
The pragmatic account has two benefits. It illuminates observed
lexical restrictions on the directional interpretation of locative
PPs, including the `squishiness' of the phenomenon in both path and
manner languages, which makes lexical diacritic approaches ultimately
infeasible. It also assumes the null hypothesis: English and Romance
are no different in their verb meanings, and, specifically, all
Romance manner of motion verbs simply lexicalize manner.
John Kingston: When does what you know affect what you hear?
What listeners know about their native language can be used to fill in content that isn’t reliably transmitted,
and it distorts the perception of foreign sounds. So there is no doubt about whether listeners’ linguistic
knowledge affects what they hear – it does –, but there is disagreement about when and how that knowledge
is applied. In interactive models (McClelland & Elman, 1986), linguistic knowledge applies to all stages in
processing as soon as it’s available, while in autonomous models (Norris, McQueen, & Cutler, 2000), only
the auditory qualities evoked by the speech signal influence its initial processing. In this talk, I present the
results of three experiments that support autonomy. The first showed that two sounds were discriminated no
better when they formed part of a word-nonword continuum than when they instead formed part of a wordword
continuum. The second showed that an early event-related potential is as robust to phonotactically
illegal [dl] followed by phonotactically legal [gl] as to phonotactically legal [dw] followed by phonotactically
legal [gw] (cf. Dupoux, Kakehi, Hirose, Pallier, & Mehler, 1999; Dehaene-Lambertz, Dupoux, &
Gout, 2000). The third showed that English listeners discriminate the input to place assimilation, [db], from
its output, [bb], far less well than a sequence which is not a possible input to place assimilation, [gb], from
its ostensible output, [bb], and likewise far less well than French listeners, whose language lacks a place
assimilation rule (cf. Darcy, Ramus, Christophe, Kinzler, & Dupoux, in press). However, this difference
disappeared when the format of the discrimination task was changed from same-different to four-interval
same-different. The results of all three experiments disconfirm the positive predictions of an interactive
model, while the results of the second and third confirm a positive prediction of the autonomous alternative,
that listeners or their brains can discriminate sounds that their knowledge renders non-distinct. The answer
to the question posed above is, “Later.”
References
Darcy, I., Ramus, F., Christophe, A., Kinzler, K., & Dupoux, E. (in press). Phonological knowledge in compensation for native and non-native assimilation. In F. K¨ugler, C. F´ery, & R. van de Vijver (Eds.),
Variation and gradience in phonetics and phonology. Berlin: Mouton De Gruyter.
Dehaene-Lambertz, G., Dupoux, E., & Gout, A. (2000). Electrophysiological correlates of phonological
processing: A cross linguistic study.
Journal of Cognitive Neuroscience, 12, 635-647.
Dupoux, E., Kakehi, K., Hirose, Y., Pallier, C., & Mehler, J. (1999). Epenthetic vowels in Japanese: A
perceptual illusion?
Journal of Experimental Psychology: Human Perception and Performance, 25,
1568-1578.
McClelland, J., & Elman, J. L. (1986). The TRACE model of speech perception.
Cognitive Psychology,
18, 1-86.
Norris, D., McQueen, J. M., & Cutler, A. (2000). Merging information in speech recognition: Feedback is
never necessary.
Behavioral and Brain Sciences, 23, 299-370.
Adam Albright: Modeling superadditive markedness interactions
A design feature of both ruled-based and constraint-based models of phonology is that processes apply independently of one another: for example, final consonant clusters with disagreeing voicing are always banned in English, causing voicing alternations regardless of whether the word has a simple or complex onset (`caps' /kæp+z/, `claps' /klæp+z/ → [kæps], [klæps]), a round vowel (`copes' /koʊp+z/ → [koʊps]), or any other marked structure. By forcing rules or constraints to apply independently, we exclude the possibility of `superadditive' effects in which the well-formedness of a structure depends on the presence of another structure. In this talk, I argue that when we move beyond alternations and turn to static phonotactics, superadditive effects do seem to occur. For example, English allows words beginning with /bl-/ and /gl-/ clusters, as well as words ending in /-sp/ and /-sk/ clusters, but there are no words with both together (
*blesk, *glisp). As it turns out, the rarity or lack of such combinations cannot be predicted from the independent frequencies of /bl-/, /-sp/, etc. I discuss several sources of evidence for superadditive effects: child errors in the acquisition of Dutch syllable structure, underattestation of doubly marked forms in the lexicons of English and Dutch, and acceptability ratings of nonce words in English. In all three cases, it appears that rare or marginal structures become worse in the presence of other rare or marginal structures. Crucially, however, not all combinations are penalized in this way. Relatively common combinations, such as /kr-/ and /-st/ co-occur about as often as expected (
crust, crest, etc.), and do not show superadditive effects.
The challenge, then, is to provide a computational model that penalizes some constraint violations more in the presence of another violation, while allowing others to remain independent. I propose a model in which acceptability judgments arise through a combination of two levels of evaluation: (1) a non-grammatical evaluation of phonotactic probability, which assesses the joint probability of the substrings in a word, and (2) evaluation by a grammar of weighted constraints, further penalizing sequences that violate high-weighted constraints. For grammatically licit combinations such as /kr-/ and /-st/, acceptability is determined by simple joint probability. For grammatically penalized clusters such as /bl-/ or /-sk/, phonotactic probability and grammatical probability combine to yield super-additive effects. I sketch a model in which learners factor out phonotactic probability in learning weights of grammatical constraints.
Jerold Edmondson: Voiced and Voiceless consonants in Taiwanese, Vietnamese, and some Formosan languages: laryngoscopic case studies
The phonetic realization of the voiced stop initials of Taiwanese Hoklo /b g/ and Vietnamese /b d/ have been described, for the first, as prenasalized but highly variable (Pan Ho-hsien, 1999) and, for the second, imploded or preglottalized (Thompson, 1977). These two language also have voiceless applosive final /p t (c) k/, which have been generally speaking, regarded as being accompanied by glottal stops with some claiming otherwise (Michaud 2004, Pham 2003). Stops in Formosan languages of Taiwan have not yet been studied in detail with the notable exception of Edmondson, Esling, Harris, Huang 2005 on Amis. This talk will report on a series of investigations (July-Dec 2009) undertaken on these consonant types employing transnasal laryngoscopic videos with full motion and frame-by-frame analysis. Our results indicate that all of these languages make use of differing lower throat structures--ventricular folds, aryepiglottic sphincter, pharyngeal- and epiglottal-elements. We will outline an analysis in terms of the "valve -model" of phonation and glottal states suggested in Edmondson & Esling 2006.
Chung-chieh Shan: Mandarin Chinese wh-indefinite scope by mixed quotation
Possible-world semantics and generalized quantifiers are two examples of systematic type-lifting. Their similarity can be codified using the notion of
monad, which simplifies the interaction between multiple levels of type-lifting in the same grammar. In particular, multiple levels of quantification let us account for the multiple readings of Mandarin Chinese sentences containing
wh-indefinites. These wh-phrases are unlike ordinary Chinese quantifiers, not only in that they require a
nonveridical licensing context for their existential force, but also in that they easily take non-surface scope:
| Haoxiang | yaoshi | wo | bu | chi | shenme, | ta | jiu | bu | zuo | de-yangzi
|
| seem | if | I | not | eat | what | he | then | not | make
|
| `It | seems | that, | if | I | don't | eat | something, | he | will | not | cook | (it).'
|
More generally, multiple levels of type-lifting arise when one language embeds another using
mixed quotation. In particular, our machinery for generating inverse scope turns out to be a special case of an account of
semantic unquotation.
Every boy claimed to like `the gift [a relative of his] gave me'.
The success of these analyses suggests that semantic rules do not float freely in a grammar but come packaged in opaque modules of linguistic competence. Specifically, monads are recursively composable modules that make semantics dynamic in the general sense of simulating pragmatics.
Ted Gibson: Quantitative investigations of syntactic representations and processes
We present a new method for quantitatively evaluating overlap in representations and processes among different linguistic structures: Inter-Subject Analysis of Co-variation (ISAC). The method is a quantitative version of a traditional syntactic approach: participants’ responses (e.g., acceptability ratings) are compared for a range of structures. This method relies on the hypothesis that a participant will rate structures that overlap in their representations or involve similar kinds of cognitive processes similarly. ISAC exploits variability present in participants’ responses to evaluate whether responses for a pair of structures co-vary systematically in a population, by means of correlations. Structures with higher overlap should produce higher correlations. We began by validating this method using wh-questions (WHQs) and relative clauses (RCs), which have long been argued to share structure. A standard observation is that WHQs and RCs are affected similarly by island constraints. We therefore applied three island constraints to WHQs and RCs: complex-NP, subject-NP, and coordination. We observed that the WHQ and RC conditions were highly correlated across different structural changes, but control structural changes (similar long-distance pronoun connections; agreement violations) did not correlate at all with the long-distance WHQ and RC conditions, suggesting that the method is potentially valuable for assessing structural and processing overlap. In a second set of experiments, we applied ISAC to pairs of constructions where the difficult version has been argued to require more working memory (WM) than the easy version: (i) nested/right-branching RCs; (ii) nested/right-branching sentential-subjects; (iii) sentential-complements embedded within RCs /RCs embedded within sentential-complements). Correlations were computed across mean difference scores for the easy vs. hard condition. Three pairs of control conditions were included: (iv) +/- NP-island-violation; (v) +/- semantic-violation; (vi) +/- agreement-violation). The results demonstrated that the WM conditions correlated more strongly with each other than with controls; the average R values differed reliably for within WM vs. between WM and non-WM (p<.01). These results suggest a shared mechanism for processing WM constructions, as predicted from the language processing literature. The initial interpretability of the results from this method across two domains suggests that it can be used to test theoretically important questions in syntax, semantics, language processing and beyond.
Liina Pylkkänen: The Neuromagnetism of Natural Language Combinatory Semantics
The essence of human language is its unbounded combinatory potential: Generative systems of syntax and semantics allow for the composition of an infinite range of expressions from a limited set of elementary building blocks. Although cognitive neuroscience has made much progress in characterizing the neural bases of lexical access, i.e., how the brain retrieves the meanings of the elementary building blocks, our understanding of the neural mechanisms of the combinatory functions remains elusive at best. A principle challenge for studying natural language combinatorics is compositionality: In many cases, the composition of complex syntactic structure is perfectly correlated with the composition of complex meaning. In an initial set of magnetoencephalography (MEG) studies, we aimed to identify a neural correlate of semantic composition by focusing on a narrow set of phenomena exhibiting a dissociation between syntactic and semantic composition. These studies consistently yielded an effect of semantic composition in a mid-line prefrontal MEG component, the Anterior Midline Field (AMF), localizing in ventromedial prefrontal cortex (vmPFC). I will discuss a series of studies aimed at further charactering the functional role of this activity as well as results implicating the left anterior temporal lobe for syntactic composition. To conclude, I will present new data assessing whether our vmPFC effect of semantic complexity is domain general, a plausible hypothesis given that the vmPFC has been implicated virtually for every domain of higher cognition.
Paolo Merlo: Joint Models of the Argument Structure of Verbs
Many models of the relationship between the syntactic and the lexical
semantic representation of verbs posit rules and representations where
these two levels interact. One expression of this interaction are the
differential frequencies of certain pairs of representations.
In this presentation, I will, first, focus on the statistical
properties that correlate the syntax and argument structure of certain
cognitively interesting classes of verbs (manner of motion, change of
state). I will show that the frequency of surface syntactic correlates
of underlying argument structure properties can be predicted based on
standard assumptions on markedness and harmonic alignment.
Statistically, the strength of this correlation is such that the
surface indicators can be used to learn the different verb classes,
for several classes and in many languages.
Second, I will use this correlation between syntactic structure and
argument assignment to argue in favor of a probabilistic parsing
model which calculates the syntactic and semantic argument structure
representations jointly. The parsing model is based on the automatic
induction of abstract representations. One same model is trained on
corpora hand annotated with syntactic and semantic role
representations for seven different languages. The results are among
the best to date.
Abby Cohn: The Internal Structure of Nasal-Stop Sequences: Evidence from Austronesian
The phonological and phonetic structure of nasal-stop sequences has elicited much attention. Yet, in fact, less is known about the internal timing of nasal-stop sequences, both unary cases (most commonly prenasalized stops–ND) and clusters (nasal voiced-stop clusters–ND, and nasal voiceless-stop clusters–NT) than often assumed. This calls into question certain phonological assumptions and conclusions. In this paper, we examine the phonetic structure of nasal-stop sequences to address this lacuna.
We present acoustic and nasal airflow data for six Austronesian languages, which between them are described to exemplify four nasal-stop sequence-types: ND, ND, ND, and NT. Erromangan and Tamambo (Oceanic languages of Vanuatu) are described as having prenasalized stops. Acehnese and Sundanese (West Austronesian languages of Indonesia) are described as having postploded nasals. Manado Malay and Pamona (also West Austronesian languages of Indonesia) are described as having nasal voiced-stop clusters. In addition, in each of these languages except for Tamambo, the nasal voiced-stop sequences contrast with nasal voiceless-stop clusters.
We will see that in each of these three types cases the central issue is timing: first, total duration is the crucial dimension that allows us to differentiate unary segments vs. clusters; second, relative timing of the nasal and oral components is critical to characterizing the difference between nasal voiced- and nasal voiceless-stop sequences; third, micro-timing (structure of the transition and nature of the oral component) is at issue in assessing whether postploded nasals are distinct from other sorts of nasal-stop sequences.
In conclusion, we find that a fuller understanding of the phonetic timing and structure of NC sequences in turn informs our understanding of phonological patterns and systems of contrast.
