The seminar will meet irregularly and will be announced on this page and via cognition@ILLC mailing list. In addition to the seminar, we have a regular reading group that meets bi-weekly. Reading group meetings will be announced via CoSaQ mailing list. If you would like to be notified of updates for that group, please contact Jakub Szymanik
To check upcoming CoSaQ seminar and reading group meetings see: CoSaQ calendar.
25.02.2021. Heming Strømholt Bremnes
Time and location: 14:00-15:30 @ Zoom
Title: Computational Complexity Explains Neural Differences in Quantifier Verification
Abstract: Quantifiers have been studied in cognitive neuroscience. What unifies the cognitive neuroscience studies, is that almost all of them implement verification paradigms. However, different classes of quantifiers require distinct verification procedures when classified in terms of their computational complexity (Szymanik 2016), and what is true of one class, is therefore not necessarily true of another. The aim of the present work is therefore to explicitly manipulate quantifier class in a verification task while recording cerebral EEG, to demonstrate that computational complexity plays a role in determining which algorithm is implemented by the brain in the verification of the different classes. Further, we are interested in at what time-points in the sentence the verification procedure diverges, and whether the effects disappear if participants perform a non-verification task with the same stimuli.
I will present behavioral and ERP data from two studies – one with explicit verification, and one without – where participants (N = 30, 27) were presented pictures of differently colored geometrical shapes, and read simple predicative sentences describing these pictures using Aristotelian, cardinal and proportional quantifiers. Divergences were found between the classes. At the noun completing the subject NP, proportional quantifiers elicited sustained positive evoked potentials relative to the other two classes, both separately and as a collapsed class, in the verification task, but not in the non-verification task. At the sentence final adjective, all sentences showed N400s for false versus true completions in the verification task, but the effects were graded with Aristotelian quantifiers showing the largest N400 and proportional quantifiers showing the smallest. In the non-verification task, N400s were found only for Aristotelian and cardinal quantifiers. Furthermore, in false trials, proportional quantifiers showed another sustained positivity relative to the other quantifiers, an effect found only in the verification experiment. These findings strongly suggest that the complexity of the algorithm required to verify different quantified expressions is reflected online in human sentence processing.
11.02.2021. Fausto Carcassi
Time and location: 14:00-15:30 @ Zoom
Title: The Shape of Modified Numerals
Abstract: The pattern of implicatures of modified numeral ‘more than n’ depends on the roundness of n. Cummins, Sauerland, and Solt (2012) present experimental evidence for the relation between roundness and implicature patterns, and propose a pragmatic account of the phenomenon. More recently, Hesse and Benz (2020) present more extensive evidence showing that implicatures also depend on the magnitude of n, argue against a pragmatic explanation, and propose a novel explanation based on the Approximate Number System (Dehaene, 1999). Despite the wealth of experimental data, no formal account has yet been proposed to characterize the full posterior distribution over numbers of a listener after hearing ‘more than n’. We develop one such account within the Rational Speech Act framework, quantitatively reconstructing the pragmatic reasoning of a rational listener. We show that our pragmatic account correctly predicts various features of the experimental data, contra Hesse and Benz (2020).
28.01.2021. Lorenzo Pinton
Time and location: 13:00-14:00 @ Zoom
Title: A few surprising data on surprisingly few
Abstract: Fernando and Kamp (1996) proposed the so-called surprise reading of `few’ and `many’. According to this reading, the two quantity words have to be conceived as `less/more than expected’. If this is the case, it is not clear if and how explicitly stating `surprisingly’ would affect the thresholds for the two quantifiers. In her dissertation, Anthea Schoeller (2017) studied the impact of context in `few’ and `many’, and in particular, in chapter 6, she tested the effects of `surprisingly’. The data she collected showed that `surprisingly’ changes the thresholds for `many’ but not for `few’, while other intensifiers (e.g. `incredibly’) change the thresholds for both. Suspecting a floor effect on `few’, we decided to run a similar experiment, implementing the proportional reading of the aforementioned quantifiers, in contrast with Schoeller’s cardinal reading. The results will be presented in Thursday’s talk.
05.11.2020. Terence Hui
Time and location: 15:00-16:30 @ Zoom
Title: Investigating the effect of context on quantifier threshold.
Abstract: In this talk, I will present the results of an experiment I ran on quantifier meaning. The experiment is based on one ran by Ramotowska et al. which tested for the thresholds of quantifiers in a neutral context. Their experiment showed that the thresholds for some quantifiers varied quite significantly between individuals. Due to the seeming variation of meaning between individuals, I hypothesized that context would have an effect on how individuals determine their thresholds for these vague quantifiers.
Thus, I ran an experiment to test whether positive or negative contexts had an effect on quantifier thresholds. To do this I ran two separate trials: one which calculated individual’s thresholds in a positive context and the other in a negative one. For certain quantifiers, there appears to be an effect of context on threshold.
22.10.2020. Iris van de Pol
Time and location: 17:00-18:30 @ Zoom
Title: Complexity of quantifiers in relation to semantic universals
Abstract: We present work in progress in which we look at the complexity of quantifiers to investigate whether the semantic universals related to monotonicity, quantity and conservativity might be explained by differences in complexity. We extend earlier work, in which we looked at approximate Kolmogorov complexity (based on Lempel-Ziv complexity) of a limited selection of quantifiers. Here, we generate a large collection of quantifiers based on a simple, yet rich, grammar. We look at two notions of complexity: shortest expression-length (based on syntax) and approximate Kolmogorov complexity (computed over the extension of an expression), and we investigate their correlation to a graded measure of the monotonicity, quantity, and conservativity of quantifiers.
08.10.2020. Nima Motamed
Time and location: 17:00-18:30 @ Zoom
Title: Quantifiers, complexity, and degrees of universals: a large-scale analysis
Abstract: It has been shown by van de Pol, Steinert-Threlkeld, and Szymanik [2019] (henceforth called vdPSS) that there is reason to believe that the differences in complexity between quantifiers’ representations might serve to explain the appearance of some semantic universals. To show this, they made use of the minimal pair methodology as also employed by Steinert-Threlkeld and Szymanik [2019] when studying the relation between semantic universals and learnability.
27.08.2019. Stefan Heim
Time and location: 15:30-17:00 @ PCH 4.11
Title: If so many are “few” for me, how few are “many” for you? Intra- and inter-individual sources of variability in quantifier processing
Abstract: While quantifiers such as “many” or “less-than-half” have a seemingly precise and distinct meaning, empirical studies show substantial variability both in use and in judgement of appropriateness. The particular external context of a situation, but also the internal reference of an individual, exert strong influence on how quantifiers are processed. I will show examples of inter- and intra-individual flexibility of quantifier usage, identify the underlying neural basis, and demonstrate how performance can systematically break down with brain disease. In the final part, I will attempt an analogy with dyslexia, referring how different individual cognitive profiles can constitute processing subtypes. The talk concludes with a discussion of which cognitive domains might be relevant determinants of quantifier usage.
07.02.2019. Fabian Schlotterbeck
Time and location: 15:30-17:00 @ PCH 5.19
Title: Decision Processes in Quantifier Verification
Abstract: Various aspects of the cognitive representation and processing of quantifier meanings can be studied by combining insights from compositional semantics with models from numerical cognition. Stipulating a transparent interface between these two cognitive subsystems, this strategy has, for example, been successfully applied to verification procedures associated with proportional quantifiers. I propose that this approach can be further complemented if we also incorporate well-established models of decision processes, so-called diffusion decision models. Thereby, we increase the amount of information to be gained from experimental data because we are able to model reaction times and proportions of errors simultaneously. Moreover, this approach promises to uncover how specific parts of linguistic representations are mapped to various components of verification processes. I discuss verification of upward and downward entailing proportional quantifiers as a concrete test case.
26.11.2018. Steven Piantadosi
Time and location: 12:00-14:00 @ PCH 3.19
Title: Statistics over algorithms as a model of human learning
Abstract: I’ll present an overview of my research that is aimed at understanding how human learners solve complex, structured learning problems. Recent theories of human learning have hypothesized that people can infer the algorithm or computation giving rise to the data they can observe. This approach shows promise in explaining human behavior across a variety of domains, including language learning and conceptual development. It also allows the field to address even more basic questions about what types of knowledge might be “built in” for humans, and how children develop the rich systems of knowledge found in adults. The ideas used in this work have the potential to inform the discovery of structure, algorithmic processes, scientific laws, and causal relations by using techniques inspired by the remarkable statistical inferences carried out by human learners. I’ll specifically talk both about computational models of number learning and rule-based concept learning, as well as experimental work aimed at testing the underlying components of those models.
22.11.2018. Fausto Carcassi
Time and location: 15:30-17:00 @ PCH 1.05.
Title: Monotonicity in gradable adjectives: an evolutionary model and some experimental results
Abstract: When it comes to conceptual domains that have a scalar structure, be it scalar adjectives, quantifiers, or modals, languages show a preference for single bounded categories, so-called monotonic categories. Most of the literature has focused on quantifiers, and shown that monotonic quantifiers are easier to learn (Steinert-Threlkeld & Szymanik forthcoming, Chemla et al. forthcoming). Two questions arise. First, the so-called problem of linkage (Kirby 1999): how does a fact about individual cognition affect language at a population level? Secondly, how to account for monotonicity in non-quantificational scalar categories, e.g. scalar adjectives?
In this talk, I discuss both questions. First, I discuss a computational model that combines Iterated Learning (Kirby et al. 2015) and Rational Speech Act modelling (Goodman & Frank 2016) to explain monotonicity in the semantics of gradable adjectives. I briefly discuss the relations between our model and previous models in the literature (Brochhagen et al 2016, 2018). Secondly, I present an experimental method that manipulates the way a set of stimuli is conceptualized to induce a preference for monotonic categories. I also present some initial (spoiler: non-significant) results from this method.
12.10.2018: Paul Pietroski
Time and location: 16:15 – 17:30 @ PCH, room 5.02
Title: Meanings, ‘Most’, and Mass
Abstract: In the talk, I will discuss a series of studies that show how experimental methods can help adjudicate between provably equivalent but procedurally distinct specifications of word meanings, focussing on ‘most’ and ‘more’. For example, while semanticists often paraphrase (1) with (2) or its formalization (2a),
(1) Most of the dots are blue.
(2) The number of blue dots exceeds the number of non-blue dots.
(2a) #{x: x is a blue dot} > #{x: x is a non-blue dot}
There is evidence that (3/3a) better represents how speakers of English understand (1).
(3) The number of blue dots exceeds the number of dots minus the number of blue dots.
(3a) #{x: x is a blue dot} > [#{x: x is a dot} – #{x: x is blue dot}]
As time permits, I’ll also review some evidence which suggests that (i) logically equivalent sentences involving ‘most’ and ‘more’ are understood in different ways, and (ii) given a single scene that can be described with a count noun like ‘dot’ or a mass noun like ‘goo’, the grammatical contrast is correlated with distinct modes of representation, in a way that tells against attempts to analyze mass nouns in terms of count nouns. In short, linguistic meanings exhibit specific representational formats that are experimentally detectable.
08.10.2018: Paul Pietroski
Time and location: 16:00 – 17:30 @ SP 107, room F1.15
Title: Confronting Existential Angst
For more information, see this page.
14.06.2018: Michael Glanzberg
Title: Binding, Compositionality, and Semantic Values (joint work with Jeffrey C. King)
For more information, see this page.
13.06.2018: Michael Glanzberg
Time and location: 15:00 – 16:30 @ PC Hoofthuis 04.28
Title: The Cognitive Roots of Adjectival Meaning
Abstract: In this paper, I illustrate a way that work in cognitive psychology can fruitfully interact with truth-conditional semantics. A widely held view takes the meanings of gradable adjectives to be measure functions, which map objects to degrees on a scale. Scales come equipped with dimensions that fix what the degrees are. Following Bartsch and Vennemann, I observe that this allows dimensions to play the role of lexical roots, that provide the distinctive contents for each lexical entry. I review evidence that the grammar provides a limited range of scale structures, presumably dense linear orderings with a limited range of topological properties. I then turn to how the content of the root can be fixed. In the verbal domain, there is evidence suggesting roots are linked to concepts. In many cases for adjectives, it is not concepts but approximate magnitude representation systems that fix root contents. However, these magnitude representation systems are approximate or analog, and do not provide precise values. I argue that the roots of adjectives like these provide a weak, discrimination-based constraint on a grammatically fixed scale structure. Other adjectives can find concepts to fix roots, which can support a well-known equivalence class construction which can fix precise values on a scale. I conclude that though adjectives have a uniform truth-conditional semantics, they show substantial differences in the cognitive sources of their root meanings. This shows that there are (at least) two sub-classes of adjectives, with roots fixed by different mechanisms and with different degrees of precision, and showing very different cognitive properties.
14.05.2018: Markus Pantsar
Title: Computational and cognitive complexity in mathematical problem solving
For more information, see this page.
09.04.2018: Peter Pagin
Title: Classical Propositions and Switcher Semantics
For more information, see this page.
16.11.2017: Galit Agmon
Time and location: 11.30-13.00, PC Hoofthuis 6.05
Title: Processing downward monotonicity in natural language.
Abstract: “A small number of circles are blue” and “few circles are blue” seem to convey the same meaning – that the quantity of blue circles is below some contextual criterion. Moreover, both contain a negative degree element (“few” as the negative of “many”; “small” as the negative of “large”). However, “few” is downward monotone while “a small number” is not. Downward Monotonicity is a formal logical property of certain linguistic expressions, but is it also cognitively relevant for language processing? In this talk I will present evidence for cognitive correlates of downward monotonicity, by presenting results of RT and fMRI experiments that compare e.g. “few” to “a small number”, and argue that the cognitive complexity of downward monotonicity is part of the mental representation of the sentence.
25.04.2017: Fabian Schlotterbeck
Time and location: 15-17, PC Hoofthuis 6.26
Title: Towards a general model of the processing difficulty of quantified sentences.
Abstract: Distinct aspects of the processing of quantifiers were largely studied separately in the past and thus a comprehensive model is lacking. For example, many psycholinguistic studies focused on Aristotelian quantifiers separately from other types of quantifiers, e.g. numerical or proportional ones. Similarly, in much recent work, processes involved in the verification of quantifiers were studied independently of other processes like, e.g., comprehension. In my second talk, I will continue the integrative approach of the first one but with a broader empirical focus. I will discuss diverse empirical findings that a comprehensive processing model of quantification would have to accommodate. Moreover, I will discuss whether and how existing theoretical models may be integrated into a general one. Among other things, two specific aspects will be discussed in some detail: The first is the relation between comprehension and verification of quantified sentences. The second is an empirical phenomenon which Bott et al. (2013, under review) have found in a range of different types of quantifiers and which they have dubbed ’empty-set effect.’
References:
Bott, O., Klein, U., & Schlotterbeck, F. (2013). Witness Sets, Polarity Reversal and the Processing of Quantified Sentences. In Proceedings of the Amsterdam Colloquium 2013 (pp. 59–66).
Bott, O., Schlotterbeck, F. & Klein, U. (under review). Empty set effects in quantifier interpretation. Submitted to Journal of Semantics
24.04.2017: Fabian Schlotterbeck
Time and location: 11-13, ILLC, F1.15
Title: Empirical evaluation of various approaches to the processing difficulty of quantified sentences and an integrated processing model of comparative modified numerals.
Abstract: One particularly interesting fact about the processing of quantifiers is that this seemingly uniform class of expressions exhibits considerable variation in how difficult they are to process. Several approaches to characterizing the processing difficulty of individual quantifiers have emerged in recent years. With regard to verification – as opposed to comprehension, for example – the most influential are (1) automata theoretic approaches, (2) logical-form based approaches and (3) approaches based on the interface transparency thesis, which posits a transparent relationship between semantic representations and the verification procedures that are actually realized within the cognitive architecture. In my talk, I will evaluate these approaches and their predictions against experimental data from the literature as well as from my own work. I will focus on two specific test cases. The first are quantified reciprocal sentences, like, e.g., ‘most dots are directly connected to each other,’ which, depending on the quantifier, may differ in computational complexity (Szymanik 2010). The second are upward vs. downward entailing comparative modified numerals, like, e.g., ‘more’ vs. ‘fewer than five,’ which are commonly observed to differ in processing difficulty (Koster-Moeller, Varvoutis & Hackl 2008). Concerning the latter, a processing model for sentence-picture verification is proposed that integrates key insights of the above-mentioned approaches. Moreover, experimental data are presented that confirm predictions of this integrated processing model.
References:
Koster-Moeller, J., Varvoutis, J. & Hackl, M. (2008). Verification procedures for modified numeral quantifiers. In Proceedings of the 27th West Coast Conference on Formal Linguistics (pp. 210–317), Somerville, MA: Cascadilla Press.
Szymanik, J. (2010). Computational complexity of polyadic lifts of generalized quantifiers in natural language, Linguistics and Philosophy, 33(3), 215–250.