University at Buffalo (UB)
Wednesday, February 2, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Stuart Shapiro, Ph.D.
Department of Computer Science and Engineering
University at Buffalo
"A Logic of Arbitrary and Indefinite Objects"
For a printable version of this file click here
Wednesday, February 9, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Selmer Bringsjord , Ph.D. & Chris Chris McEvoy
Department of Cognitive Science
Artificial Intelligence and Reasoning Laboratory
Rensselaer Polytechnic Institute
"Building a Virtual Person (E) from the "Dark Side"
We describe our general approach to building what we call advanced synthetic characters (or *bona fide* virtual persons), within the paradigm of logic-based AI. This approach, based on our RASCALS architecture, seeks to use a cognitive architecture for ``mid-level" cognition, and advanced logical systems for more advanced reasoning-intensive thought. To focus our general approach, we provide a glimpse of our attempt to bring to life one particular advanced synthetic character from the "dark side" --- the character known simply as E (for, as you may have guessed, evil). Building E entails, among other things, that we formulate an underlying logico-mathematical definition of evil, and that we manage to engineer as well an appropriate presentation of E.
At the presentation level, we use an approach based in manipulating facial musculature.
For a printable version of this file click here
Wednesday, February 16, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Conor McLennan , Ph.D.
Department of Psychology
Language Perception Laboratory
University at Buffalo
"Variability in Spoken Word Recognition"
For a printable version of this file click here
Wednesday, February 23, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Jürgen Bohnemeyer , Ph.D.
Department of Linguistics
University at Buffalo
"Manner and path in nonlinguistic cognition"
(Joint research with Sonja Eisenbeiss (University of Essex)
and Bhuvana Narasimhan (Max Planck Institute for Psycholinguistics)
In this paper, we present findings from a wide crosslinguistic survey, designed to investigate whether language-specific patterns of motion event encoding along the lines of Talmy’s (1985, 2000) typology of verb-framed vs. satellite-framed languages (1, 2 below) influence nonlinguistic cognition.
1. Verb-framed languages:
The ball entered the cave rolling PATH encoded in main verb, MANNER in dependent
2. Satellite-framed languages:
The ball rolled into the cave MANNER encoded in main verb, PATH in dependent
Slobin (1996) and Berman & Slobin (1994) suggest that differences in linguistic event descriptions can result in differences in “thinking for speaking”. Gennari et al. (2002) and Papafragou et al. (2002) found that performance in nonlinguistic categorization tasks does not reflect language-specific influences, although Gennari et al.'s study reveals an effect of prior verbal encoding of the motion event. Finkbeiner et al. (2002) found a language-specific effect on similarity judgments only in case memory recall was involved in the task. However, these studies only pit speakers of two languages against each other (three in the case of Finkbeiner et al.), and variables such as manner and path are treated as monolithic concepts, ignoring finer-grained distinctions (e.g. path expressions vary in terms of directionality, boundary crossing, etc.; manner expressions vary in whether they imply translational motion (e.g., slide, walk vs. spin, bounce), whether the motion is self-propelled (walk vs. slide), etc.).
In order to further investigate these issues, we conducted a nonlinguistic similarity judgment task which systematically varies types of manners and paths in a range of typologically diverse languages (12 V-framed languages: Basque, Catalan, Hindi, Italian, Jalonke, Japanese, Lao, Spanish, Tamil, Turkish, Tiriyo, Yukatek; 3 S-framed languages: Tidore, Dutch, German). Twelve native speakers of each language viewed a target motion event (e.g. ball rolling up a ramp) followed by two events which varied from the target in its manner of motion (e.g. ball sliding up the ramp) or path of motion (e.g. ball rolling down the ramp). Participants judged which of the two variants was more similar to the target. It was hypothesized that speakers of S-framed languages would prefer the event which had the same manner of motion as the target (even though the path of motion is different).
Our findings reveal a significant effect of language. However, the effect is not based on the S-framed versus V-framed distinction. Rather, we find intra-typological variation. V-framed languages fall into two groups, one whose speakers strongly prefer to categorize the stimuli on the basis of manner of motion, and one whose speakers show a weak preference for categorization by path. Speakers of S-framed languages do not differ significantly from either group. Further, there are significant effects of finer-grained contrasts in path and manner. The observed effects of path type are language-independent: triads which involved a vertical (up-down) path elicited a significantly lower manner preference overall than triads with a horizontal (left-right) path. The effects of particular manner contrasts, however, vary according to language: for instance, Spanish speakers are more likely than German speakers to accept a sliding display as a variant of a rolling display, whereas German speakers are more likely than Spanish speakers to accept a sliding display as a variant of a spinning display. The implications of our findings, their relation to existing work on these issues, and lines of future research will be discussed.
For a printable version of this file click here
Wednesday, March 2, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
William Rapaport , Ph.D.
Department of Computer Science and Engineering
Department of Philosophy and
Center for Cognitive Science
University at Buffalo
"In Defense of Contextual Vocabulary Acquisition:
How to Do Things with Words in Context"
"Context" is notoriously vague, and its uses multifarious. Researchers in "contextual vocabulary acquisition" differ over the kinds of context involved in vocabulary learning, and the methods and benefits thereof.
This talk presents a computational theory of contextual vocabulary acquisition, identifies the relevant notion of context, exhibits the assumptions behind some classic objections, and defends our theory against these objections.
References:
Beck, Isabel L.; McKeown, Margaret G.; & McCaslin, Ellen S. (1983),
"Vocabulary Development: All Contexts Are Not Created Equal",
Elementary School Journal 83(3): 177-181.
http://ublib.buffalo.edu/cgi-bin/reserve.cgi?B029442831.PDF
Schatz, Elinore Kress, & Baldwin, R. Scott (1986),
"Context Clues Are Unreliable Predictors of Word Meanings",
Reading Research Quarterly 21(4, Fall): 439-453.
http://ublib.buffalo.edu/cgi-bin/reserve.cgi?B029441932.PDF
Rapaport, William J. (submitted, 2004),
"In Defense of Contextual Vocabulary Acquisition:
How to Do Things with Words in Context",
submitted to Context-05.
http://www.cse.buffalo.edu/~rapaport/Papers/paris.pdf
For a printable version of this file click here
Wednesday, March 23, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Elsi Kaiser , Ph.D.
Center for Language Sciences
University of Rochester
"Picture of who? An experimental investigation of pronouns and reflexives in representational noun phrases"
The observation that English pronouns and reflexives have a (nearly) complementary distribution is central to standard binding theory (BT). Representational NPs (RNPs, e.g. 'picture of her/herself') are a well-known exception, as both pronouns and reflexives are acceptable (e.g. Kuno 1987, Pollard & Sag 1992, Reinhart & Reuland 1993, Tenny 2003). Thus, they may provide a useful window into the syntax/pragmatics/semantics interface. In this talk I discuss experiments we conducted on English and Finnish investigating (i) the idea that reflexives in RNPs refer to "sources-of-information" (see Kuno 1987) and (ii) Tenny's observation that pronouns in RNPs refer to "perceivers-of-information." We used the action-based visual-world paradigm, which crucially provides both time-course data and information about the referent assigned to the anaphor on each trial. The results for English picture NP constructions show that both reflexives and pronouns are influenced by source and perceiver information respectively, but that the effects are much stronger for pronouns and, crucially, arise even when binding theory is not violated. The results for Finnish, a typologically different language with greater morphological complexity, show that in postnominal RNP constructions, there is a perceiver preference for pronouns as well as a source preference for certain reflexive forms. Thus, in the Finnish reflexive system, morphological differences correspond to interpretational differences. However, prenominal RNP constructions in Finnish show no source/perceiver effects for either anaphoric option. In sum, on the basis of the English data we can conclude that discourse/semantic factors interact with BT, but affect pronouns with local antecedents more than reflexives with non-local antecedents. The Finnish data suggest that whether discourse/semantic factors interact with BT depends on the structural domain, since different domains show presence and absence of verb effects in RNPs. As a whole, the findings suggest that in order to better understand the referential properties of pronouns and reflexives, we need to take into account not only the structural configuration but also other kinds of information such as the source/perceiver distinction.
For a printable version of this file click here
Wednesday, March 30, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Dan Gildea , Ph.D.
Computer Science Department
University of Rochester
Syntactic Structure and Statistical Machine Translation
Given that statistical methods have revolutionized both natural language parsing and machine translation, it may seem surprising that most current statistically-based translation systems make no use of syntactic structure.
I will describe work on models of translation that aim to fill this gap, presenting results for models that make use of syntactic information provided for one or both languages, as well as models that infer structure directly from parallel bilingual text. I will also describe the use of syntactic information for the automatic evaluation of machine-produced translations.
For a printable version of this file click here
Wednesday, April 6, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Julia Hirschberg , Ph.D.
Computer Science Department
Columbia University
"Recognizing a Speaker's Emotional State"
A speaker's emotional state is conveyed by acoustic and prosodic factors, as well as the words they choose and the gestures they use. We are studying several different contexts in which emotional state is important to determine: 1) an automatic tutoring system, in which students studying physics may be confident or uncertain, frustrated, or angry, and should receive appropriate handling for that state; 2) speech in varied public settings, where speakers may be perceived as charismatic or not, providing some indication of the likely success of speakers' attempts to gain political power; and 3) recorded interviews in which speakers may be telling the truth or not. In each case, our focus is on identifying prosodic and acoustic as well as lexical cues to these different speaker states, so that we may develop systems which automatically distinguish between, e.g., confidence and uncertainty, frustration and satisfaction, charimatic and non-charismatic speech, and deceptive and non-deceptive speech. These studies represent joint work with the University of Pittsburgh, SRI International, and the University of Colorado.
For a printable version of this file click here
Wednesday, April 20, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Len Talmy , Ph.D.
Department of Linguistics
University at Buffalo
"The Attention System of Language"
This talk reports on work in progress to outline the fundamental attentional system of language. This system includes some fifty basic factors, the "building blocks" of the system. Each factor involves a particular linguistic mechanism that increases or decreases attention on a certain type of linguistic entity. Although able to act alone, the basic factors also regularly combine and interact to produce further attentional effects. This attentional system shows commonalities and differences across individual languages, across modalities (spoken vs. signed language), and across cognitive systems (e.g., between language and visual perception). The methodology used in the analysis, introspection, is itself made the subject of investigation to determine its profile of better and worse function and its consequent relation to other methodologies.
For a printable version of this file click here
Wednesday, April 27, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Dedre Gentner , Ph.D.
Department of Psychology
Cognitive Science Program
Department of Education and Social Policy
Northwestern University
"Acquiring and Using Relational Representations: Computational and Empirical Details and Theoretical Speculations"
For a printable version of this file click here
3:30 pm - 5:00 pm
Baird Concert Hall, North Campus
Dedre Gentner , Ph.D.
Department of Psychology
Cognitive Science Program
Department of Education and Social Policy
Northwestern University
"Why we're so smart"
Human cognitive abilities are remarkable, and even more remarkable is the the rapidity with which children develop cognitive insight. How does this insight arise? A pervasive view in cognitive development is that these rapid gains can only be explained by assuming that infants begin with substantial amounts of innate knowledge. In this talk I propose an alternative approach, centered on mechanisms of human learning. I suggest two powerful forces that contribute to human learning and reasoning ability: (1) analogical processing; and (2) the acquisition of relational language. I will present evidence that the structure-mapping processes that occur during analogy and similarity are a core mechanism by which abstract knowledge arises from experience. Our studies of learning in adults and children show that analogical comparison processes foster learning in several ways: by aligning common relational structure, by suggesting inferences between situations, by focusing attention on relevant differences, and by inviting relational abstractions.
A further contributor to human learning and reasoning is the acquisition of relational language. Relational language provides labels that preserve and systematize the relations discovered through comparison processes. It also acts to invite analogical comparisons that reveal common structure. In sum, I suggest that mutual bootstrapping between structure-mapping processes and relational language is a major contributor to human cognition.
About Dedre Gentner:
Dedre Gentner’s research is on the psychology of learning and reasoning and the development of cognition and language. Her early work on causal mental models and on the development of word meaning have been influential in cognitive research. Her most important contribution is the structure-mapping theory of analogy and similarity and its implications, including a computational model of similarity processing; a theoretical framework for analogy and metaphor; the evidence for disassociation between the kind of similarity that governs memory retrieval and the kind of similarity that governs on-line mapping and inference. In her developmental work she has proposed a relational shift in children’s similarity processing and has found evidence that this shift is knowledge-driven, rather than maturational. She has also proposed and tested a progressive alignment mechanism whereby comparison processes in ordinary experience can yield theoretical insight.
In language learning, Gentner’s hypothesis of a language-universal advantage for nouns in children’s early word learning that has engendered considerable research. Her recent work unites analogical thinking and language learning and investigates possible interactions between language and cognition. Her theoretical and empirical work provides evidence that relational language has a formative role in the development of relational thought. She is also investigating the hypothesis that analogical
processes are integral to language acquisition and use.
Wednesday, September 14, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Mary Swift, Ph.D.
Computer Science Department
University of Rochester
"Default aspect and the development of
time reference in Inuktitut child language"
Children acquiring the temporal system of Inuktitut, a polysynthetic language spoken by the Inuit of Arctic Quebec, exhibit developmental phenomena that appear puzzling in comparison to previous crosslinguistic findings. Inuit children very early on use a single unmarked verb form for two kinds of time reference: perfective with telic event descriptions, and imperfective with atelic event descriptions. In contrast, previous research shows that across languages children in their early speech use different tense-aspect marking with telic event descriptions than they use with atelic event descriptions. The first puzzle is how children acquiring Inuktitut come to terms with the variable time reference of the unmarked verb form, without local cues to facilitate differentiation of interpretation. Second, Inuit children's early instances of past and perfective marking occur with atelic event descriptions, in contrast to previous findings showing that initial instances of past and/or perfective marking crosslinguistically occur with telic event descriptions, a pattern predicted by e.g., Basic Child Grammar (Slobin 1985) and The Aspect Hypothesis (Shirai and Anderson 1995). Most strikingly, Inuit children develop facility with overt future marking before overt past marking, a pattern that has not been reported for any other language. These three puzzles of Inuktitut temporal language development find a uniform account under the analysis of Bohnemeyer and Swift (2004), which is based on the notion of event realization (ER). Informally, ER is the factual occurrence of an event at a certain time. ER plays a double role in the analysis of Inuktitut child language: in Inuktitut, the temporally unmarked verb form receives its aspectual interpretation under ER; I also show that aspectual reference in early Inuktitut is constrained by ER, and that this constraint applies to early child language crosslinguistically.
Sponsored by the Department of Linguistics, University at Buffalo
For a printable version of this file click here
Wednesday, September 21, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Amnon Eden, Ph.D.
Department of Computer Science
University of Essex
U.K.
"Software Ontology as a Cognitive Artefact"
This talk is concerned with the philosophy of software. I will discuss views in the metaphysics of software and why we should care about it. I will examine the merits of viewing software ontology as a cognitive artefact and draw out some of the insights concerning the elaborate dynamics between programming languages and programming paradigms. The clams advanced are illustrated using programs in a variety of programming languages. In conclusion I will present a conceptual scheme that brings together all entities relating to software, from meta-software to hardware: the mind-to-machine taxonomy.
Amnon H Eden, Department of Computer Science, University of Essex, United Kingdom, and Center For Inquiry, Amherst, NY.
Bio: http://www.eden-study.org/bio.html
Sponsored by the Department of Philosophy, University at Buffalo
For a printable version of this file click here
Wednesday, September 28, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Ron Smyth , Ph.D.
Division of Humanities and Division of Life Sciences
University of Toronto
"Sociophonetics, Gender, and Sexual Orientation"
This talk with give an overview of the sociophonetics research I have been conducting with Henry Rogers for the past few years at the University of Toronto. In this work we extend the restrictive notion of gender as typically found in phonetics and sociolinguistics -- namely, the distinction between speakers who are biologically male vs. female -- to considerations of sexual identity. This involves comparing the voices of gay/lesbian speakers to those of straight speakers, and, perhaps more importantly, examining how acoustic and articulatory parameters correlate with how gay/lesbian a voice is judged to be, regardless of the actual sexual orientation of the speaker. These variables include mean pitch, intonational variability, sibilant duration and frequency, /l/ velarization, voice onset time, vowel duration, formant structure, and (coming soon) voice quality characteristics. In addition to these correlational findings we have also done multivariate analysis to identify which properties are intercorrelated vs. those that seem to be independently available to speakers (Arnold Zwicky's "Menu Hypothesis").
Deeper questions involve how and why sexual identity affects speech, the extent to which 'gay' speech is 'female' speech, how and when gendered speech is acquired, and both the advantages and the limitations of laboratory-based research as compared to more subtle sociolinguistic interview methodologies.
Sponsored by the Department of Linguistics, University at Buffalo
For a printable version of this file click here
Wednesday, October 12, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Change in speaker, please watch for the new announcement
For a printable version of this file click here
Wednesday, October 19, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Werner Ceusters , MD.
European Centre for Ontological Research
Saarland University, Saarbrücken, Germany
"Computational Linguistics for Referent Tracking in Electronic Healthcare Records: a research agenda"
In the last two decades we have witnessed considerable efforts directed towards making electronic healthcare records (EHRs) comparable and interoperable through advances in record architectures and (bio)medical terminologies and coding systems. Deep semantic issues in general, and ontology in particular, have received some interest from the research communities. However, with the exception of work on so-called ‘controlled vocabularies’, ontology has thus far played little role in work on standardization. The prime focus has been rather the rapid population of terminologies at the level of fine detail with the purpose of making them available for “clinical coding”. The rationale for clinical coding is the belief that codes will make it possible to associate precise meanings to the terms used in expressing patient data in a way that can be interpreted by software for further processing for purposes such as statistic analysis, billing, reimbursement, automated decision support, and triggering alerts. However, many aberrations in the biomedical coding and classification systems and terminologies that are used today, prevent such further processing to be done in a reliable way. This is because the terms or codes contained in the latter are used simply as an alternative to what would otherwise have been registered by means of general terms in natural language. By picking a code from such a system and then registering that code in an EHR, one refers generically to some instance of the class represented by the code. It is still left at best only partially specified which particular instance (“referent”) is intended in concrete reality.
It is here that referent tracking comes into play. The goal of referent tracking is to create an ever-growing pool of data relating to concrete entities in reality. In the context of Electronic Healthcare Records (EHRs) the relevant concrete entities are not only particular patients but also their body parts, diseases, therapies, lesions, and so forth insofar as these are relevant to their diagnosis and treatment. Within a referent tracking system, all such entities are referred to explicitly, something which cannot be achieved when only the codes from concept-based systems are used.
A particular challenge for the referent tracking paradigm in the context of EHRs is to minimize the amount of overhead that users would experience when entering data in the records. Ideally, they should be able to continue to work in the same way as before, either by writing clinical narrative or by working with biomedical terminologies, while software running in the background should replace generic codes with entity identifiers where applicable. This can be viewed as a modified version of referent resolution which is known to be a very hard problem in computational linguistics.
About Werner Ceusters:
Werner Ceusters studied medicine ('77-'84) neuropsychiatry ('84-90'), informatics ('88-'90) and knowledge engineering ('91-'93). He started a series of international research projects in medical natural language processing under the Third, Fourth and Fifth Research Frameworks of the European Commission through his R&D company Office Line Engineering nv. Since then, he has also been active in standardisation bodies related to medical terminology such as CEN/TC251/WG2 and ISO/TC215/WG3. In April 1998, he started a new company - Language & Computing nv (L&C) - to exploit the results of his research. He left L&C in 2004 and created together with Prof. Barry Smith the European Centre for Ontological Research, his main interest being now applying and testing a new theoretically-grounded approach to ontological engineering.
References:
Sponsored by the Department of Philosophy, University at Buffalo
For a printable version of this file click here
Wednesday, October 26, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Boris Hennig , Ph.D.
Institute for Formal Ontology
Saarland University
Germany
"Normativity and the Mental"
What we call mental event, Descartes calls thought. This is how he defines it:
By the term ‘thought’, I understand everything that happens in us of which we are conscious, insofar as we are conscious of it.
Descartes tells us that events that happen in us are thoughts insofar as they are the object of our consciousness. What is consciousness? It seems to be a certain way of perceiving something that happens in us. However, the problem with such an account is that perceiving something would itself be a kind of thought. This would mean that Descartes circularly defines thought as the object of a special kind of thought.
I will maintain that consciousness as it appears in this definition is not a further mental event but a general normative stance. Events that happen in us are thus mental events by virtue of being subject to certain norms. That they are subject to these norms means that they may be, but not necessarily that they are evaluated according to them. The normative stance need not be actually taken by anyone. Hence, consciousness need not happen in us. If this is correct, then consciousness is not a mental event.
Sponsored by the Department of Philosophy, University at Buffalo
For a printable version of this file click here
Wednesday, November 2, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Tsan Huang , Ph.D.
Department of Linguistics
University at Buffalo
"Effects of linguistics experience
on early levels of tone perception"
This study investigated the phenomenon of language-specificity in Mandarin Chinese tone perception. The main question was whether linguistic experience affects the earliest levels of perceptual processing of tones. Chinese and American English listeners participated in four perception experiments, which involved short inter-stimulus intervals (300ms or 100ms) and an AX discrimination or AX degree-of-difference rating task. Three experiments used natural speech monosyllabic tone stimuli and one experiment used time-varying sinusoidal simulations of Mandarin tones. AE listeners showed psychoacoustic listening in all experiments, paying much attention to onset and offset pitch. Chinese listeners showed language-specific patterns in all experiments to various degrees, where tonal neutralization rules reduced the perceptual distance between two otherwise contrastive tones for Chinese listeners. Since these experiments employed procedures hypothesized to tap the auditory trace mode, language-specificity found in this study seems to support the proposal of an auditory cortical map (Guenther et al. 1999). But the model needs refining to account for different degrees of language-specificity, which are better handled by Johnson's (2004) lexical distance model, although the latter model is too rigid in assuming that linguistic experience does not affect low-level perceptual tasks such as AX discrimination with short ISIs.
For a printable version of this file click here
Wednesday, November 9, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Cassandre Creswell
Janya, Inc., Buffalo, NY
"Automatically detecting nominal mentions of events"
The basic task in information extraction (IE) is to automatically extract mentions of entities, relations, and events from text. Usually, events are treated as textually anchored by verb phrases or sentences, as in (1), but of course they can also be referred to with noun phrases, as in (2).
1) [Amr Moussa was appointed Secretary General of the Arab League last year.]
2) [Amr Moussa's appointment as Secretary General of the Arab League] took place last year.
Automatic detection of nominal mentions of events faces several challenges. The one we address here is the systematic polysemy unique to nominal (vs. verbal) mentions, such that they can be used to denote either an event or the result, outcome, or product of an event.
3) Things are getting back to normal in the Baywood Golf Club after [a chemical spill](=event). Clean-up crews said [the chemical spill](=result) was 99 percent water and shouldn't cause harm to area residents.
This ambiguity is problematic for IE systems because the ultimate goal in event extraction is to create abstract event objects containing information from multiple coreferring mentions, such as their time, location, and participants. Information should not be propagated from mentions of non-events to events, and event mentions should not be collapsed with non-event mentions.
Tackling this ambiguity requires more fine-grained syntactic and semantic evidence than is used in standard word sense disambiguation techniques (Schuetze, 1998) because the event-result ambiguity does not obey the One Sense Per Discourse principle (Gale, et al., 1992).
The goal of this project is to develop a classifier that can label NPs as event-denoting or non-event-denoting based on their local lexico-syntacticcontext. Because learning from hand-annotated data alone is not sufficient to overcome the sparse data problem, we use a weakly-supervised bootstrapping technique using known unambiguous terms to learn to distinguish the contexts which make an event vs. non-event interpretation more probable. In this talk I will present some initial experimental results from our ongoing project on nominal event detection.
Sponsored by the Department of Linguistics, UB
For a printable version of this file click here
Wednesday, November 30, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Deb Roy, Ph.D.
Massachusetts Institute of Technology
Department of Media Arts and Sciences
Cognitive Machines Group, MIT Media Laboratory
"Meaning Machines"
The meaning of words in everyday situated language use depends on physical and social context. Our group is building robots and computer game players that use situated language as a way to study and model the physical and social grounding of semantics. Emerging from these implementations are conceptual representations underlying word meanings based on semiotic schemas, perceived affordances, and spatial routines. We are using these representations and associated learning algorithms to develop computational models of language acquisition that “step into the shoes” of children and learn directly from what children hear and see. I will provide an overview of our research program and highlight our recently launched Human Speechome Project which is poised to collect and model an unprecedented level of in-home audio-visual observational data of child language development from birth to age three.
Sponsored by the Department of Computer Science and Engineering, UB
For a printable version of this file click here
Wednesday, December 7, 2005
2:00 pm - 4:00 pm
280 Park Hall, North Campus
Lou Ann Gerken, Ph.D.
Department of Psychology
Cognitive Science Program
University of Arizona
Sponsored by the Departments of Psychology and Linguistics.
Some Generalizations About Linguistic
Generalization By Infants
One dimension on which more vs. less strongly constrained models of language acquisition vary is the amount of evidence required for a particular linguistic generalization. ‘Triggering’ models require, in the limit, only a single datum to set an innate parameter, whereas less constrained models often arrive at a generalization by performing statistics over many exemplars from an input set. I will present data from three lines of research with 9- to 17-month-old infants, which explore the amount and type of input required for learners to generalize beyond the stimuli encountered in a brief laboratory exposure. All of the studies suggest that generalization requires a minimal number of data points, but more than just one, and that different subsets of the input lead to different generalizations. Some surprising new data suggest that some types of generalization (those attested by natural languages) may be easier to make than others (those unattested by natural languages). Taken together, the data provide direction for examining the ways in which innate constraints and learning via statistics may combine in human language development.
Sponsored by the Departments of Linguistics and Psychology.