Research

2021. Task influences on lexical underspecification: Insights from the Maze and SPR, coauthored with Jack Duff and Amanda Rysling ++ abstract, slides
To be presented at CUNY 2021.

In assigning an incremental interpretation to linguistic input, sometimes a resolution to a temporary ambiguity will prove incompatible with downstream material, leading to costly reanalysis. However, in some cases comprehenders appear to avoid this cost. One approach to such observations is to posit that comprehenders can underspecify some input, delaying full commitment and interpretation. One set of data often taken as evidence for underspecification is Frazier & Rayner (1990)’s landmark eyetracking investigation of polysemy. However, it remains unclear why polyseme specification is delayed: is the underspecification utility-based (used under some comprehension strategies but not others); or is it necessary due to some property of lexical representation and semantic commitment? We address this question by replicating Frazier & Rayner (1990) across additional tasks, Maze and SPR. We observe task-dependent variation in line with the utility account: polyseme underspecification is present in SPR but is avoided in Maze, a fact we attribute to the task’s demands for eager interpretation.

2021. The representation and processing of distributivity and collectivity: ambiguity vs. underspecification, coauthored with Jakub Dotlačil. Glossa 6(1):6, 1–22 ++ paper

We introduce a framework in which production-rule based computational cognitive modeling and Reinforcement Learning (RL) can systematically interact and inform each other. We focus on linguistic applications because of the complex systems needed to capture linguistic cognitive behavior. On the cognitive modeling side, we can start exploring in a computationally explicit way how complex, mechanistically specified, and production-based cognitive models of linguistic skills, e.g., ACT-R based parsers, can be acquired. Linguistic cognitive model learnability is a largely understudied issue, and RL can help us investigate it in a formally and computationally explicit way. In turn, the sophisticated rule-based cognitive models needed to capture human behavioral data provide a very stringent test suite for RL algorithms – more so than video games in some respects, and with the added advantage that RL algorithms can be explicitly connected to experimental data (both forced-choice and reaction-time). We can thus assemble an experimentally rigorous and cognitively realistic benchmark for RL algorithms.

2020. Production-based Cognitive Models as a Test Suite for Reinforcement Learning Algorithms, coauthored with Jakub Dotlačil ++ paper, poster
To be presented at CMCL 2020.

We introduce a framework in which production-rule based computational cognitive modeling and Reinforcement Learning (RL) can systematically interact and inform each other. We focus on linguistic applications because of the complex systems needed to capture linguistic cognitive behavior. On the cognitive modeling side, we can start exploring in a computationally explicit way how complex, mechanistically specified, and production-based cognitive models of linguistic skills, e.g., ACT-R based parsers, can be acquired. Linguistic cognitive model learnability is a largely understudied issue, and RL can help us investigate it in a formally and computationally explicit way. In turn, the sophisticated rule-based cognitive models needed to capture human behavioral data provide a very stringent test suite for RL algorithms – more so than video games in some respects, and with the added advantage that RL algorithms can be explicitly connected to experimental data (both forced-choice and reaction-time). We can thus assemble an experimentally rigorous and cognitively realistic benchmark for RL algorithms.

2020. Reinforcement Learning for Production-based Cognitive Models, coauthored with Jakub Dotlačil ++ abstract, slides, video
Talk at AMLaP 2020, Special session: Computational models of language processing.

We introduce a framework in which we can start exploring in a computationally explicit way how complex, mechanistically specified, and production-based cognitive models of linguistic skills, e.g., ACT-R based parsers, can be acquired. Cognitive models for linguistic skills pose this learnability problem much more starkly than models for other ‘high-level’ cognitive processes, as they call for richly structured representations and complex rules that require a significant amount of hand-coding. In this paper, we focus on how Reinforcement Learning (RL) methods can be used as a way to solve the production selection problem in ACT-R. Production rules are treated as the actions of an RL agent, and the ACT-R model/mind as the environment in the RL sense. We focus on a basic learning algorithm (tabular Q-learning) and a very simple task, namely lexical decision (LD), framed as a sequential decision problem, with the goal of learning when a specific rule should be fired. Learning is faster and less noisy for shorter LD tasks (fewer stimuli), but the Q-learning agent manages to learn longer tasks fairly well. Realistically long LD tasks and more complex models, e.g., parsers, are left for future research.

2020. Effects of morphological identity and voice mismatch in VP ellipsis, coauthored with Jesse Harris ++ abstract, poster
Presented at AMLaP 2020.

We investigate asymmetrical Voice Mismatch effects in VP ellipsis, in which Passive-Active mismatches (John was praised by Mary, and Sue did, too) are reportedly tolerated better than Active-Passive ones. We propose that Passive forms passively activate form-identical lexical representations for Active forms, creating the illusion of a matching antecedent. We make the novel prediction that Passives with morphologically distinct forms from Actives (was driven~drove) do not activate Active counterparts, and are thus less susceptible to interference from lexical competition than those with identical forms (was found~found). An acceptability-rating task on a 7-point Likert scale and a speeded 2AFC acceptability study confirms our predictions.

2020. Semantic commitment and lexical underspecification in the Maze, coauthored with Jack Duff and Amanda Rysling ++ abstract, poster
Presented at AMLaP 2020.

A complete model of incremental sentence processing must encompass how and when lexical ambiguities are resolved in real time. In a landmark study, Frazier & Rayner (1990) demonstrated in eyetracking while reading that such resolution depends on the nature of the ambiguity, contrasting homonymy and polysemy. The present study seeks to replicate these results with a different reading measure, the Maze task (Forster, Guerrera & Elliot 2009). The Maze requires strictly incremental comprehension at each decision point, enabling us to investigate how susceptible underspecification of polysemy is to task pressures. Results from this study thus suggest that underspecification of polysemy cannot be an architectural necessity: underspecification seems to be an optional heuristic, foregone in the Maze due to atypical task pressures.

2020. Textual Evidence for the Perfunctoriness of Independent Medical Reviews, coauthored with Megan Moodie and Rakshit Agrawal ++ paper, slides, video
Presented at KDD Workshop on Knowledge-infused Mining and Learning (KiML).

We examine a database of 26,361 Independent Medical Reviews (IMRs) for privately insured patients, handled by the California Department of Managed Health Care (DMHC) through a private contractor. IMR processes are meant to provide protection for patients whose doctors prescribe treatments that are denied by their health insurance (either private insurance or the insurance that is part of their worker comp; we focus on private insurance here). Laws requiring IMR were established in California and other states because patients and their doctors were concerned that health insurance plans deny coverage for medically necessary services. We analyze the text of the reviews and compare them closely with a sample of 50000 Yelp reviews [19] and the corpus of 50000 IMDB movie reviews [10]. Despite the fact that the IMDB corpus is twice as large as the IMR corpus, and the Yelp sample contains almost twice as many reviews, we can construct a very good language model for the IMR corpus using inductive sequential transfer learning, specifically ULMFiT [8], as measured by the quality of text generation, as well as low perplexity (11.86) and high categorical accuracy (0.53) on unseen test data, compared to the larger Yelp and IMDB corpora (perplexity: 40.3 and 37, respectively; accuracy: 0.29 and 0.39). We see similar trends in topic models [17] and classification models predicting binary IMR outcomes and binarized sentiment for Yelp and IMDB reviews. We also examine four other corpora (drug reviews [6], data science job postings [9], legal case summaries [5] and cooking recipes [11]) to show that the IMR results are not typical for specialized-register corpora. These results indicate that movie and restaurant reviews exhibit a much larger variety, more contentful discussion, and greater attention to detail compared to IMR reviews, which points to the possibility that a crucial consumer protection mandated by law fails a sizeable class of highly vulnerable patients.

2020. Reinforcement Learning for Production-based Cognitive Models, coauthored with Jakub Dotlačil ++ paper, slides, video
Presented at MathPsych/ICCM 2020.

We introduce a framework in which we can start exploring in a computationally explicit way how complex, mechanistically specified, and production-based cognitive models of linguistic skills, e.g., ACT-R based parsers, can be acquired. Cognitive models for linguistic skills pose this learnability problem much more starkly than models for other ‘high-level’ cognitive processes, as they call for richly structured representations and complex rules that require a significant amount of hand-coding. In this paper, we focus on how Reinforcement Learning (RL) methods can be used as a way to solve the production selection problem in ACT-R. Production rules are treated as the actions of an RL agent, and the ACT-R model/mind as the environment in the RL sense. We focus on a basic learning algorithm (tabular Q-learning) and a very simple task, namely lexical decision (LD), framed as a sequential decision problem, with the goal of learning when a specific rule should be fired. Learning is faster and less noisy for shorter LD tasks (fewer stimuli), but the Q-learning agent manages to learn longer tasks fairly well. Realistically long LD tasks and more complex models, e.g., parsers, are left for future research.

2020. Reinforcement Learning for Production-based Cognitive Models, coauthored with Jakub Dotlačil ++ poster
Presented at CUNY 2020.

We introduce a framework in which we can start exploring in a computationally explicit way how complex, mechanistically specified, and production-based cognitive models of linguistic skills (e.g., ACT-R based parsers) can be acquired. Linguistic cognitive model learnability is a largely understudied issue, primarily because computationally explicit cognitive models are only starting to be more widely used in psycholinguistics. Cognitive models for linguistic skills pose this learnability problem much more starkly than models for other ‘high-level’ cognitive processes, since cognitive models that use theoretically-grounded linguistic representations and processes call for richly structured representations and complex rules that require a significant amount of hand-coding.

2020. Cognitive Modeling for Formal Semantics, coauthored with Jakub Dotlačil ++ poster
Presented at CUNY 2020.

We outline the structure of a cognitively realistic semantic processor: a parser that can incrementally construct semantic representations. The constructed semantic representations are Discourse Representation Structures (DRS, see Kamp 1981; Kamp and Reyle 1993), similar to mental models in the sense of Johnson-Laird (1983, 2004). We show the workings of the semantic parser by accounting for the reaction time (RT) data in the fan experiment reported in Anderson (1974). This is the first (to our knowledge) fully implemented and cognitively realistic model of incremental interpretation and truth evaluation that is systematically informed by formal semantics and can model RT data.

2020.The processing of direct discourse: When a subordinate speech act sticks around, coauthored with Jack Duff, Pranav Anand and Amanda Rysling ++ poster
Presented at CUNY 2020.

Direct discourse (DD) speech reports contribute information which is foreign to the current discourse, a distinction shown to be relevant at early stages of reading. Yet, over two naturalness-judgement studies, we demonstrate that late-stage processing does not reflect any secondary status for DD. These results are challenging for a hypothesis derived from work on appositive relative clauses (ARCs), that all peripheral discourse units are downweighted during late-stage processing. Further work must revisit whether and how the processor is sensitive to distinctions at the level of discourse structure.

2020. Morphological identity and voice mismatch in VP ellipsis: An interference-based account, coauthored with Jesse Harris ++ poster
Presented at CUNY 2020.

We investigate asymmetrical Voice Mismatch effects in VP ellipsis, in which Passive-Active mismatches (John was praised by Mary, and Sue did, too) are reportedly tolerated better than Active-Passive ones. We propose that Passive forms passively activate form-identical lexical representations for Active forms, creating the illusion of a matching antecedent. We make the novel prediction that Passives with morphologically distinct forms from Actives (was driven~drove) do not activate Active counterparts, and are thus less susceptible to interference from lexical competition than those with identical forms (was found~found). A speeded 2AFC acceptability study confirms our predictions.

2020. Computational Cognitive Modeling and Linguistic Theory OPEN ACCESS LINK, book in the Springer Language, Cognition, and Mind (LCAM) series (vol. 6), coauthored with Jakub Dotlačil

The two blurbs below are also available on the Springer page for the book:

This open access book introduces a general framework that allows natural language researchers to enhance existing competence theories with fully specified performance and processing components. Gradually developing increasingly complex and cognitively realistic competence-performance models, it provides running code for these models and shows how to fit them to real-time experimental data. This computational cognitive modeling approach opens up exciting new directions for research in formal semantics, and linguistics more generally, and offers new ways of (re)connecting semantics and the broader field of cognitive science.

The approach of this book is novel in more ways than one. Assuming the mental architecture and procedural modalities of Anderson’s ACT-R framework, it presents fine-grained computational models of human language processing tasks which make detailed quantitative predictions that can be checked against the results of self-paced reading and other psycho-linguistic experiments. All models are presented as computer programs that readers can run on their own computer and on inputs of their choice, thereby learning to design, program and run their own models. But even for readers who won’t do all that, the book will show how such detailed, quantitatively predicting modeling of linguistic processes is possible. A methodological breakthrough and a must for anyone concerned about the future of linguistics! (Hans Kamp)

This book constitutes a major step forward in linguistics and psycholinguistics. It constitutes a unique synthesis of several different research traditions: computational models of psycholinguistic processes, and formal models of semantics and discourse processing. The work also introduces a sophisticated python-based software environment for modeling linguistic processes. This book has the potential to revolutionize not only formal models of linguistics, but also models of language processing more generally. (Shravan Vasishth)

Summary:

The book integrates formal semantic theories, specifically, DRT, and mechanistic processing models formulated in the ACT-R cognitive architecture into a new computational framework that can simulate human participants in real-time psycholinguistic tasks, fit these cognitive models to experimental data by embedding them into Bayesian models, and quantitatively compare qualitative (symbolic) theories.

In more detail, we introduce an extensible framework for mechanistic processing models that integrate formal syntax and semantics theories, an independently motivated cognitive architecture (ACT-R; Anderson and Lebiere 1998, Anderson 2007) and Bayesian methods of data analysis, parameter estimation and model comparison. The book culminates with the development of computationally-explicit processing models for natural language semantic phenomena that integrate Discourse Representation Theory (DRT; Kamp 1981, Kamp & Reyle 1993), ACT-R and Bayesian models.

These mechanistic processing models implement end-to-end simulations of human participants in real-time psycholinguistic experiments. End-to-end means that we model visual and motor processes together with specifically linguistic representations and processes in complete models of the experimental tasks. The models embed theoretical hypotheses about linguistic representations and about parsing processes in ACT-R. The resulting ACT-R cognitive models are then embedded in Bayesian models to fit them to experimental data, estimate their parameters and perform quantitative model comparison for qualitative (symbolic) theories.

A novel Python3 reimplementation of ACT-R called pyactr (current version; book version, including all book code) provides the crucial software infrastructure for developing ACT-R models that can be easily embedded in Bayesian models.

2020. Reinforcement learning for production-based cognitive models, joint work with Jakub Dotlačil
UCSC S-circle talk, May 29, 2020.

2020. Language models and sentiment classification
UCSC AAII Tutorial Series, May 22, 2020.

2020. Deep learning for collaborative filtering and tabular modeling
UCSC AAII Tutorial Series, May 18, 2020.

2020. Binary and multiclass image classification
UCSC AAII Tutorial Series, May 15, 2020.

Building a prototype image classification project, from data collection to web deployment The anatomy, training and evaluation of a simple digit classifier Image classification: A more realistic example

2019. Cognitive Modeling for Formal Semantics: The Organization of DRSs in Declarative Memory, coauthored with Jakub Dotlačil ++ paper, abstract, very short demo of the semantic parser, longer demo of the semantic parser, poster, lightning talk slides
In the Proceedings of the Amsterdam Colloquium 2019.

One of the key and still outstanding challenges for formal semantics is “how to build formal semantics into real-time processing models – whether psychological or computational – that involve the integration of linguistic and not-specifically linguistic knowledge.” (Partee 2011) In this paper, we outline the structure of a cognitively realistic semantic processor that we have fully implemented elsewhere (Brasoveanu and Dotlačil 2020). The processor is able to incrementally construct DRT semantic representations in response to linguistic stimuli. We show how this semantic parser can be used to account for the reaction time (RT) data in the fan experiment reported in Anderson (1974), which investigates how propositional information of the kind encoded by atomic DRSs is (stored and) retrieved from declarative memory. We provide the first (to our knowledge) fully implemented and cognitively realistic model of incremental interpre- tation and semantic evaluation that is systematically informed by formal semantics and can be used to model RT data.

2019. Transfer Learning in NLP
UCSC AAII Tutorial Series, November 2019

slides (pdf), slides (pptx)

text preprocessing (html), text preprocessing (ipynb)

the AWD-LSTM model (html), the AWD-LSTM model (ipynb)

pretraining on Wikipedia (html), pretraining on Wikipedia (ipynb)

ULMFiT transfer learning (html), ULMFiT transfer learning (ipynb)

2019. Practical Bayesian Inference ++ slides (html), slides (ipynb)
UCSC AAII Tutorial Series, November 2019

2019 Computational Cognitive Modeling for Syntax and Semantics
University of Connecticut Logic Group presentation, April 2019.

2018. Cognitively realistic computational models for natural language understanding, coauthored with Jakub Dotlačil
Santa Cruz Machine Learning Group presentation, UCSC, October 2018.

We introduce a typical experimental task in psycholinguistics — self-paced reading — and show how to build end-to-end simulations of a human participant in such an experiment; end-to-end means that we model visual and motor processes together with specifically linguistic processes in a complete model of the experimental task. The model embeds theoretical hypotheses about linguistic representations and parsing processes in an independently motivated cognitive architecture (ACT-R). In turn, the resulting cognitive models can be embedded in Bayesian models to fit them to experimental data, estimate their parameters and perform quantitative model comparison for qualitative theories.

2018. The semantics and processing of correlatives ++ slides
Presented at the Multiple wh-constructions and their kin workshop, Nantes, September 2018, organizers: Anamaria Falaus and Ivano Caponigro.

2018. An extensible framework for mechanistic processing models: From representational syntax-semantics theories to quantitative model comparison, coauthored with Jakub Dotlačil ++ paper, slides
Presented at MathPsych/ICCM 2018.

We introduce a Python3 reimplementation of ACT-R (Anderson and Lebiere 1998, Anderson 2007) in which we build an end-to-end simulation of syntactic parsing in a typical self-paced reading experiment. The model uses an eager left-corner parsing strategy implemented as a skill in procedural memory (following Lewis and Vasishth 2005), makes use of independently motivated components of the ACT-R framework (procedural memory, content-addressable declarative memory – cf. Wagers et al. 2009), and explicitly models the motor and visual processes involved in self-paced reading. The ACT- R model can be embedded in a Bayesian statistical model to estimate its subsymbolic parameters and perform model comparison.

2018. A cognitively realistic left-corner parser with visual and motor interfaces: An extensible framework for mechanistic processing models, coauthored with Jakub Dotlačil ++ abstract, slides
Presented at CUNY 2018; also presented at UCSC S-Lab (March 2018).

We introduce a Python3 implementation of ACT-R (Anderson and Lebiere 1998) in which we build an end-to-end simulation of syntactic parsing in a typical self-paced reading experiment. The model uses a left-corner parsing strategy implemented as a skill in procedural memory (following Lewis and Vasishth 2005), makes use of independently motivated components of the ACT-R framework (content-addressable declarative memory, Wagers and Phillips 2009, goal and imaginal buffers etc.), and explicitly models the motor and visual processes involved in self-paced reading. The ACT-R model can be embedded in a (Bayesian) statistical model to estimate its sub-symbolic parameters and do model comparison.

2018. Modeling lexical access in ACT-R, coauthored with Jakub Dotlačil ++ abstract, poster
Presented at CUNY 2018.

One robust parameter affecting latencies and accuracies in lexical decision tasks is frequency. Since Howes and Solomon (1951), it is accepted that lexical access can be approximated as a log-function of frequency. In their Exp. 1, Murray and Forster (2004) (M&F) collected responses and response times in a lexical decision task using words from 16 frequency bands, and showed that log-frequency provides an imperfect fit to the data. We provide an ACT-R model of the M&F data and embed it in a Bayesian model to estimate its parameters. The results cast doubt on some common assumptions in ACT-R psycholinguistic models.

2018. Quantitative Comparison for Generative Theories: Embedding Competence Linguistic Theories in Cognitive Architectures and Bayesian Models, coauthored with Jakub Dotlačil ++ abstract, slides, handout (slides 4up), paper
Presented at BLS 2018; also presented at UCSC S-Lab (January 2018).

We introduce a new framework in which (i) qualitative generative grammar hypotheses can be embedded into performance/processing models, and (ii) these integrated competence-performance models are embedded into Bayesian models, which enables us to fit them to experimental data and do quantitative model comparison for qualitative theories. Building on Lewis and Vasishth (2005), step (i) takes generative theories and embeds them in processing theories formulated in the ACT-R cognitive architecture (Anderson and Lebiere 1998). In a new Python3 computational implementation of ACT-R, we do step (ii), namely use ACT-R models as the likelihood component of full Bayesian models. We demonstrate (i) and (ii) by comparing models for lexical decision tasks that differ qualitatively and/or quantitatively.

2018. Expressing ignorance in Japanese: contrastive wa vs. sukunakutomo, coauthored with Hitomi Hirayama. Accepted in final form, Journal of Cognitive Science
Presented at the 2017 East Asian Psycholinguistics (EAP): Recent developments workshop (held in conjunction with the 25th Japanese Korean Linguistics conference)

2018. An Experimental Investigation of the Scope of Object Comparative Quantifier Phrases, coauthored with Kristen Syrett & Atsushi Oho. Accepted in final form, Journal of Semantics

2018. Processing Quantification, coauthored with Jakub Dotlačil. In press, Handbook of Experimental Semantics and Pragmatics. Chris Cummins & Napoleon Katsos (eds.), Oxford University Press

2018. Kinds of (Non-)Specificity, coauthored with Donka Farkas. To appear in the semantics portion of Wiley’s Linguistics Companion (Companion to Semantics, SemCom). D. Gutzmann, L. Matthewson, C. Meier, H. Rullmann & T.E. Zimmermann (eds.), Wiley

2018. Donkey Anaphora: Farmers and Bishops, coauthored with Jakub Dotlačil. To appear in the semantics portion of Wiley’s Linguistics Companion (Companion to Semantics, SemCom). D. Gutzmann, L. Matthewson, C. Meier, H. Rullmann & T.E. Zimmermann (eds.), Wiley

2017. Evaluativity Across Adjective and Construction Types: An Experimental Study, coauthored with Jessica Rett. In Journal of Linguistics, 1-67

An adjectival construction is ‘evaluative’ if and only if the property associated with the adjective is assumed to be instantiated to a significantly high degree. We report a series of experiments that target evaluativity (norm-relatedness) of gradable adjectives of different types across several degree constructions. Existing theoretical studies of evaluativity – e.g. that John is as tall as Sue is not evaluative, while John is as short as Sue is – are largely based on introspective, informally collected judgments. The main contribution of this paper is a fairly rich set of experimentally-supported generalizations that constitute a first step towards a more solid empirical foundation for the existing theoretical literature on evaluativity. In particular, our experiments substantiate prior claims that evaluativity is conditioned by adjective type (relative or absolute, Kennedy and McNally 2005). They also confirm prior claims that evaluativity is not restricted to the positive construction (Rett 2008b), although our results imply that evaluativity is best viewed as a gradable property. We additionally report on interesting discrepancies between the received generalizations and our experimental results. First, the relative/absolute distinction affects evaluativity differently across construction types: for positive constructions, absolute adjectives are more evaluative than relative adjectives, but this pattern is reversed in the case of equatives. Second, although the existing literature characterizes comparatives as not evaluative (e.g., John is taller than Sue is felicitous even if John does not count as tall), our Experiments 1 and 2 show that the subject position as clearly more evaluative than the ‘object’ (comparison term), and than either position in equatives. In a third follow-up experiment, this result is explained by the observation that participants seem to use the descriptive content of the adjectival constructions themselves to determine a standard for the purposes of evaluativity (in a way reminiscent of van Rooij 2011). This complicates the claim in Barker 2002 that evaluative constructions receive either a descriptive or metalinguistic interpretation, and has consequences for experimental studies and semantic theories of adjectival evaluativity, and context-sensitive phenomena more generally.

2016. Dynamic Semantics, coauthored with Jan van Eijck, Rick Nouwen & Albert Visser. Entry in the Stanford Encyclopedia of Philosophy

2016. The Semantics of Corrections, coauthored with Deniz Rudin, Karl DeVries, Karen Duek and Kelsey Kraus. WoLLIC 2016, LNCS 9803. J. Väänänen et al. (eds.), 358-374

2015. Incremental and Predictive Interpretation: Experimental Evidence and Possible Accounts, coauthored with Jakub Dotlačil ++ abstract, slides, handout (slides 4up), paper
Presented at SALT 25, UCSC S-circle (April 2015) and Georgetown Linguistics Speaker Series (April 2015).

The main question we will address in this talk is whether meaning representations of the kind that are pervasive in formal semantics are built up incrementally and predictively when language is used in real time, in much the same way that the real-time construction of syntactic representations has been argued to be (Steedman 2001, Lewis and Vasishth 2005, Lau 2009, Hale 2011 among many others).

The interaction of presupposition resolution with conjunctions vs. conditionals with a sentence-final antecedent promises to provide us with the right kind of evidence. Consider the contrast between the ‘cataphoric’ examples in (1) and (2) below:

(1) Tina will have coffee with Alex again AND she had coffee with him at the local café.

(2) Tina will have coffee with Alex again IF she had coffee with him at the local café.

We expect (1) to be less acceptable / more difficult than (2) right after the presupposition trigger “again” is interpreted: the conjunction “and” signals that an antecedent that could resolve the “again” presupposition is unlikely to come after this point (the second conjunct is interpreted relative to the context provided by the first conjunct), while the conditional “if” leaves open the possibility that a suitable resolution for the “again” presupposition is forthcoming (the first clause is interpreted relative to the context provided by the second clause). Crucially, the different expectations triggered by the interaction of the presupposition trigger “again” and the operators “and” vs. “if” are semantically driven: there is nothing about the syntax of conjunction vs. “if”-adjunction that could make the presence of a suitable presupposition resolution more or less likely.

2015. Processing Pluralities: Syntax and the Lexicon, coauthored with Jakub Dotlačil ++ abstract, poster
Presented at CUNY 28.

In semantics, two types of distributivity/collectivity are standardly distinguished: lexical and phrasal. In the former case, the distributivity/collectivity targets a word and could be seen as a purely lexical phenomenon: e.g., in the boys won, the interpretations “{each boy / the group of boys} won” can be derived just from the vagueness of won, i.e., from the fact that won can be true of groups or individuals. Phrasal distributivity cannot be specified in the lexicon: in the boys won an award, the interpretation that each boy won a different award requires us to assume that the whole predicate (verb + indefinite) is interpreted distributively when composed with the subject. Phrasal distributivity is derived with the help of a dist operator requiring the combination of a predicate and its subject to be interpreted distributively (Winter 2000, Kratzer 2013, a.o.). The distinction between phrasal and lexical distributivity/collectivity, known in semantics, was ignored in previous psycholinguistic research. Our study is the first to take it into account. Our experimental findings are compatible with the position that the processor disprefers phrasal distributivity. Distributivity per se is not dispreferred since no general cost of post-verbal distributive disambiguation was seen. The results provide evidence for two distributivity types in grammar. They are compatible with the hypothesis that only phrasal distributivity requires a distributivity operator in syntax (under the assumption that the human parser prefers minimal syntactic structures — Minimal Attachment, Frazier 1978, a.o.).

2015. Kinds of (Non-)Specificity, coauthored with Donka Farkas.
In preparation for the semantics portion of Wiley’s Linguistics Companion (Companion to Semantics, SemCom). Lisa Matthewson, Cécile Meier, Hotze Rullmann, Thomas Ede Zimmermann (eds.).

2015. Donkey Anaphora: Farmers and Bishops, coauthored with Jakub Dotlačil
In preparation for the semantics portion of Wiley’s Linguistics Companion (Companion to Semantics, SemCom). Lisa Matthewson, Cécile Meier, Hotze Rullmann, Thomas Ede Zimmermann (eds.).

2015. Dynamic Semantics, coauthored with Jan van Eijck, Rick Nouwen, Albert Visser
In preparation. This is a revised version of the current Stanford Encyclopedia of Philosophy entry on Dynamic Semantics.

2015. Scales of Negativity, coauthored with Donka Farkas and Floris Roelofsen ++ paper (coming soon)
In preparation.

Abstract coming soon.

2015. The Polysemy of Container Pseudo-partitives, coauthored with Karen Duek ++ paper
In the Proceedings of Sinn und Bedeutung 19.

A pseudo-partitive phrase headed by a container word, such as bucket of water, may be interpreted as either the container object or its contents. In “Mary broke the bucket of water”, bucket of water satisfies the selectional restriction of break if it is interpreted as a container, while in “Mary spilled the bucket of water” spill requires it to be the amount of water. Existing approaches (Selkirk 1977, Rothstein 2009, Partee and Borschev 2012) take the two readings to arise from different syntactic and semantic representations. This paper tests this hypothesis experimentally using acceptability judgments and the co-predication test. We asked if the two interpretations of a container construction are available concurrently in the same representation, and can therefore satisfy the contradictory selectional requirements of two predicates in the same sentence as in “Mary broke the bucket of water that Chris spilled”. Our results show that unlike the prediction of current theories, co-predication does not lead to ungrammaticality. We propose instead an analysis of the container construction based on Asher’s (2011) dot-type framework that treats the pseudo-partitive as a complex object constructor.

2015. Strategies for Scope Taking, coauthored with Jakub Dotlačil ++ paper
In Natural Language Semantics 23:1, 1-19.

This squib reports the results of two experimental studies, a binary choice and a self-paced reading study, that provide strong support for the hypothesis in Tunstall (1998) that the distinct scopal properties of each and every are at least to some extent the consequence of an event-differentiation requirement contributed by each (Experiment 3 in Tunstall 1998, Chapter 5 designed to provide such evidence does not reveal the predicted effect). However, we also show that the emerging picture is more complex than the one painted in Tunstall (1998): English speakers seem to fall into at least three groups regarding the scopal properties of each and every.

2015. Sentence-internal Same and Its Quantificational Licensors: A New Window into the Processing of Inverse Scope, coauthored with Jakub Dotlačil ++ paper
In Semantics and Pragmatics 8:1, 1-52.

This paper investigates the processing of sentence-internal same with four licensors (all, each, every and the) in two orders: licensor+same (surface scope) and same+licensor (inverse scope). Our two self-paced reading studies show that there is no general effect of surface vs. inverse scope, which we take as an argument for a model-oriented view of the processing cost of inverse scope: the inverse scope of quantifiers seems to be costly because of model structure reanalysis, not because of covert scope operations. The second result is methodological: the psycholinguistic investigation of semantic phenomena like quantifiers and the licensing of sentence-internal readings should always involve a context that prompts a deep enough processing of the target expressions. In one of our two studies, participants read the target sentences after reading a scenario introducing the two sets of entities the quantifier NP and the same NP referred to and they were asked to determine whether the sentence was true or false relative to the background scenario every time. In the other study, the participants read the same sentences without any context and there were fewer follow-up comprehension questions. The relevant effects observed in the study with contexts completely disappeared in the out-of-context study, although the participants in both studies were monitored for their level of attention to the experimental task.

2015. The Manner and Time Course of Updating Quantifier Scope Representations in Discourse, coauthored with Jakub Dotlačil ++ paper
In Language, Cognition and Neuroscience (formerly Language and Cognitive Processes), 30:3, 305-323.

We present the results of two experiments, an eye-tracking study and a follow-up self-paced reading study, investigating the interpretation of quantifier scope in sentences with three quantifiers: two indefinites in subject and object positions and a universal distributive quantifier in adjunct position. In addition to the fact that such three-way scope interactions have not been experimentally investigated before, they enable us to distinguish between different theories of quantifier scope interpretation in ways that are not possible when only simpler, two-way interactions are considered. The experiments show that contrary to underspecification theories of scope, a totally ordered scope-hierarchy representation is maintained and modified across sentences and this scope representation cannot be reduced to the truth- conditional/mental model representation of sentential meaning. The experiments also show that the processor uses scope-disambiguating information as early as possible to (re)analyze scope representation.

2014. Ignorance in Context: The Interaction of Modified Numerals and QUDs, coauthored with Matthijs Westera ++ abstract, poster, paper
To appear in the Proceedings of SALT 24.

We argue for a purely pragmatic account of the ignorance inferences associated with superlative but not comparative modifiers (at least vs. more than). Ignorance inferences for both modifiers are triggered when the question under discussion (QUD) requires an exact answer, but when these modifiers are used out of the blue the QUD is implicitly reconstructed based on the way these modifiers are typically used, and on the fact that at least n, but not more than n, mentions and does not exclude the lower bound exactly n. The paper presents new experimental evidence for the context-sensitivity of ignorance inferences, and also for the hypothesis that the higher processing cost reported in the literature for superlative modifiers is context-dependent in the exact same way.

2014. Predicting Strict vs. Sloppy Reflexives in VP ellipsis, coauthored with Matthew Ong ++ abstract, paper
In Empirical Issues in Syntax and Semantics 10, the book associated with the CSSP 2013 conference.

The VP ellipsis (VPE) ‘remnants’ did too/did in sentences like (1) and (2) below can have two different readings depending on whether the reflexive pronoun himself covertly recovered at the VPE site is interpreted as referring back to John (the strict reading) or to Bill (the sloppy reading).

1. John blamed himself and Bill did too. [Bill blamed: John (strict)/Bill (sloppy)]
2. John blamed himself because Bill did. [Bill blamed: John (strict)/Bill (sloppy)]

Generally both strict and sloppy readings are available in (2) while the strict reading in (1) is marginal or inaccessible. Hestvik 1995’s approach to VPE argues that this asymmetry arises because of the difference in syntactic configurations between the two clauses (parallel vs. subordinate). In contrast, Kehler 2002’s discourse-based account of VPE argues the asymmetry derives from a difference in rhetorical relations (resemblance vs. cause-effect). Partial experimental support for the syntactic approach is provided by Experiment 1 in Kim and Runner 2009. This paper reports a series of two experimental studies with two related goals. The first goal is to empirically evaluate the syntactic and discourse accounts, building and improving on the experimental investigation in Kim and Runner 2009. The results strongly suggest that discourse-based accounts are empirically more adequate. The second goal is to argue that a heretofore ignored lexical factor, namely the meaning of the ellided verb, is in fact a strong predictor of strict vs sloppy readings. We found that so-called implicit-causality verbs which are object-oriented are much more likely to have strict readings than subject-oriented implicit-causality verbs.

2014. Question Tags and Sentential Negativity, coauthored with Karen De Clercq, Donka Farkas, Floris Roelofsen ++ paper
A revised version of this ms. appeared in Lingua.

This paper presents an experiment that is designed to quantify the negativity of sentences with different types of negative operators (n-words like never and downward operators like rarely) in different syntactic positions (adverb, subject, and direct object). In the experiment, participants were asked to form a tag-question based on a given sentence, adding either a positive or a negative tag. Clearly positive sentences (i.e., sentences without any negative operators) and clearly negative sentences (i.e., sentences with overt sentential negation) were used as controls. The relative frequency of positive and negative tags was then taken as a measure of the sentential negativity of each experimental item. Our main finding is that sentential negativity is not a binary, categorical notion, contrary to what is commonly assumed, but rather a graded notion, sensitive to both semantic and syntactic factors. Semantically, we find that n-words contribute more negativity than downward entailing operators. Syntactically, we find that negative items in subject or adverbial position contribute more negativity that negative items in direct object position.

2013. What a Rational Interpreter Would Do: Building, Ranking, and Updating Quantifier Scope Representations in Discourse, coauthored with Jakub Dotlačil ++ slides, handout, paper
Invited talk at the Quantitative Data in Formal Semantics and Pragmatics workshop, 19th Amsterdam Colloquium.

We begin by framing the general problem of ‘rationally’ (in the sense of Anderson et al’s ACT-R framework) integrating formal semantic theories and processing and indicate how the relation between semantic theories and the processor could be explicitly formalized. An explicit formalization would enable us to empirically evaluate the integrated theory of semantics and processing both qualitatively and quantitatively. We then introduce the problem of quantifier scope, in particular the processing difficulty of inverse scope, introduce two types of theories of scope and their predictions, and discuss the results of an eye-tracking and a self-paced reading experiment and their consequences for the two types of theories of scope. We end by outlining how probabilities for LF construction rules could be computed based on the experimental results, thereby enabling our theories to make quantitative, not only qualitative, predictions.

2013. Presuppositional Too, Postsuppositional Too, coauthored with Anna Szabolcsi ++ paper
In The dynamic, inquisitive, and visionary life of φ, ?φ, and ◊φ: a festschrift for Jeroen Groenendijk, Martin Stokhof, and Frank Veltman. Maria Aloni, Michael Franke, and Floris Roelofsen (eds.), ILLC Publications, 55-64.

One of the insights of dynamic semantics in its various guises (Kamp 1981, Heim 1982, Groenendijk & Stokhof 1991, Kamp & Reyle 1993 among many others) is that interpretation is sensitive to left-to-right order. Is order sensitivity, particularly the default left-to-right order of evaluation, a property of particular meanings of certain lexical items (e.g., dynamically interpreted conjunction) or is it a more general feature of meaning composition? If it is a more general feature of meaning composition, is it a processing ‘preference’ or should it be captured as a ‘harder’ constraint on the type of meanings and operations over meanings involved in natural language interpretation? This squib draws attention to the symmetrical A-too B-too construction (found in a variety of languages, e.g., Hungarian, Japanese, Romanian, Russian) in this context. It argues that any semantic analysis of its main ‘symmetrical-meaning’ characteristic should also allow for subtler interactions between this construction and items that are clearly sensitive to evaluation-order effects, e.g., anaphoric adjectives like next and other. We suggest that the notion of postsupposition embedded in a broader dynamic framework is better able to account for both the symmetric nature of this construction, its non-symmetric variant A-too, and its interaction with items that are evaluation-order sensitive. We briefly compare this proposal with a couple of possible alternative accounts.

2013. A Typology of Specificity, coauthored with Donka Farkas
In Revue Roumaine de Linguistique, issue associated with the Workshop on the Syntax and Interpretation of Specificity (Bucharest University 2012).

The paper casts a look back at the role specificity has played in DP semantics in the last thirty years and then defends the proposal that the common thread across specificity distinctions is the contrast between stability vs. variability in value assignments for the variable introduced by the DP across various types of alternatives. Special determiners are used to mark either variation (non-specificity) or stability (specificity). We exemplify with a discussion of the singular determiner some in English.

2013. Indefinites, coauthored with Donka Farkas
Chapter for the Cambridge Univ. Press Handbook of Semantics, M. Aloni (ed.).

2013. At-issue Proposals and Appositive Impositions in Discourse, coauthored with Scott AnderBois and Robert Henderson ++ paper
A revised version of this ms. appeared in Journal of Semantics.

Potts (2005) and many subsequent works have argued that the semantic content of appositive (non-restrictive) relative clauses, e.g., the boldfaced material in John, who nearly killed a woman with his car, visited her in the hospital, must be in some way separate from the content of the rest of the sentence, i.e., from at-issue content. At the same time, there is mounting evidence from various anaphoric processes that the two kinds of content must be integrated into a single, incrementally evolving semantic representation. The challenge is how to reconcile this informational separation with these pervasive anaphoric connections. We propose a dynamic semantic account that accomplishes this by taking appositive and at-issue content to involve two different kinds of updates to the Context Set (CS). Treating the context set as a distinguished propositional variable, we argue that appositives directly impose their content on the CS variable by eliminating possible values assigned to it. In contrast, we treat at-issue assertions as introducing a new propositional dref and proposing that the CS variable be updated with its content, subject to addressee’s response. In addition to capturing the behavior of appositives in discourse, we show that the account can be extended to capture the projection of appositive content past various sentential operators.

2013. N-words and Sentential Negation: Evidence from polarity particles and VP ellipsis, coauthored with Donka Farkas and Floris Roelofsen ++ paper
In Semantics and Pragmatics 6, 7:1-33.

Sentences involving n-words, such as No student stepped forward, have been treated either as involving sentential negation taking the n-word in its scope or as involving a negative quantifier occurring in an otherwise positive sentence. This paper provides novel experimental evidence for the former view based on data involving polarity particles and VP ellipsis, offering new tools for detecting the presence of sentential negation.

2013. Modified Numerals as Post-suppositions ++ paper
A revised version of this ms. appeared in Journal of Semantics.

The paper provides a compositional account of cumulative readings with non-increasing modified numerals (a.k.a. van Benthem’s puzzle), e.g., Exactly 3 boys saw exactly 5 movies. The main proposal is that modified numerals make two kinds of semantic contributions. Their asserted / at-issue contribution is a maximization operator that introduces the maximal set of entities that satisfies their restrictor and nuclear scope. The second contribution is a post-supposition, i.e., a cardinality constraint that needs to be satisfied relative to the context that results after the at-issue meaning is evaluated. Thus, the interpretation process ends up giving a kind of pseudowide scope to the post-suppositions contributed by modified numerals.

2013. The Grammar of Quantification and the Fine Structure of Interpretation Contexts ++ paper
A revised version of this ms. appeared in Synthese.

Providing a compositional interpretation procedure for discourses in which descriptions of complex dependencies between interrelated objects are incrementally built is a key challenge for formal theories of natural language interpretation. This paper examines several phenomena and argues that we need richly structured contexts of interpretation that are passed on between different parts of the same sentence and also across sentential boundaries to account for these phenomena. The main contribution of the paper is showing how we can add structure to contexts in an incremental way, starting with the basic notion of context in classical first-orderlogic, i.e., interpretation contexts formalized as single total variable assignments.

2012. The Online Interpretation of Sentence Internal Same and Distributivity, coauthored with Jakub Dotlačil ++ abstract, slides, handout, paper
In the Proceedings of SALT 22.

This paper investigates how sentence-internal same is processed with three of its licensors (each, all and the) and two orders: licensor+same, i.e., surface scope (Each student saw the same movie) and same+licensor, i.e., inverse scope (The same student saw each movie). Our study shows that (i) there is no effect of surface vs inverse order for all, which we take as an argument for a model-oriented view of the processing cost of inverse scope, and (ii) each and the are processed slower than all, which we take as an argument for a particular semantics of these distributive licensors and sentence-internal same.

2012. The Semantics and Processing of Distributivity, coauthored with Jakub Dotlačil ++ slides

This ms. reports the results and theoretical consequences of 3 eye-tracking experiments that investigated the following issues in the semantics and processing of distributivity: (i) whether distributive vs. collective readings for plural definites are a matter of ambiguity or vagueness (building on previous work by Frazier et al 1999) — and if the contrast is a matter of ambiguity, which reading is the default one (if any); (ii) the contrast between semantic and pragmatic biases in the interpretation of plural definites; (iii) the contrast between quantificational / scopal distributivity (e.g., the floating quantifier each) and lexical distributivity (e.g., tall is a lexically distributive predicate).

2012. Licensing Sentence-internal Readings in English: An Experimental Study, coauthored with Jakub Dotlačil ++ abstract, poster, paper
In the Proceedings of Amsterdam Colloquium 18.

Adjectives of comparison (AOCs) like same, different and similar can compare two elements sentence-internally, i.e., without referring to any previously introduced element. This reading is licensed only if a semantically plural NP is present. We argue in this paper that it is incorrect to describe a particular NP as either licensing or not licensing the sentence-internal reading of a specific AOC: licensing is more fine-grained. We use experimental methods to establish which NPs license which AOCs and to what extent and we show how the results can be interpreted against the background of a formal semantic analysis of AOCs. Finally, we argue that using Bayesian methods to analyze this kind of data has an advantage over the more traditional, frequentist approach.

2012. The pragmatics of quantifier scope: a corpus study, coauthored with Scott AnderBois and Robert Henderson ++ abstract, slides, handout, paper
In the Proceedings of Sinn und Bedeutung 16; also presented at UCLA (Linguistics Dpt. Colloq., fall 2011), UCB (S-circle, winter 2012), The Proper Use of Quantification in Ordinary Language workshop (ESSLLI 2011), MIT Syntax-Square, May 9 2011 and CUSP 3, Stanford, Oct. 15-16 2010.

Most investigations of quantifier scope are concerned with the range of possible scopes for sentences with multiple quantifiers. Instead, this study examines the actual scopes (i.e., the pragmatics of quantifier scope disambiguation) in a naturally quantifier-rich corpus: LSAT Logic Puzzles. The three main findings of our investigation are as follows. First, we confirmed findings in previous literature that linear order and grammatical function have an effect on scope-taking preferences. Second, we discovered that lexical effects on scoping preferences are at least as important as linear order or grammatical function. Third, the relational aspect of these lexical effects, i.e., the lexical realizations of the other quantifiers in the sentence, is also important. The present investigation opens the way towards a broader research program of identifying scoping-behavior patterns that should ultimately enable us to group quantifiers into classes depending on the type of scopal behavior they exhibit. These classes could provide an empirical basis for semantic theories that assign different kinds of semantic representations to different classes and / or for psycholinguistic theories that hypothesize different processing strategies for different classes.

2012. Polarity particles and the anatomy of n-words, coauthored with Donka Farkas and Floris Roelofsen ++ abstract, handout, paper
In the Proceedings of Sinn und Bedeutung 16.

Sentences involving n-words, like “None of the men stepped forward”, have been treated either as involving an indefinite expression within the scope of sentential negation (the NI approach), or as involving negative quantifiers occurring in otherwise positive sentences (the NQ approach). This paper provides novel experimental evidence for the NI approach based on data involving polarity particles in English, and thereby offers a new tool for diagnosing sentence level negation.

2012. Correlatives ++ paper
A revised version of this ms. appeared in Language and Linguistics Compass.

The paper reviews the semantics and pragmatics of correlatives across various ontological domains (temporal, modal, individual and degree domains), focusing mostly on Indo-European languages. Taking correlatives in the individual domain as a case study, the paper then argues that their interpretation, in particular the variability of the uniqueness effects exhibited by correlatives, is due to their mixed referential and quantificational nature. The account involves an articulated notion of quantification consisting of both (discourse) referential components and non-referential/quantificational components - thus bringing together previous analyses that took either the referential or the quantificational route. The variable uniqueness effects emerge out of the interaction between (i) the semantics of wh-indefinites, singular anaphors and habitual morphology and (ii) the pragmatics of quantification, which allows for the selection of different levels of ‘zoom-in’ on the quantified-over objects.

2011. Tutorial on dynamic semantics and talk on the semantics of sentence-internal readings of different et co., Stanford University, part of the Context Dependence workshop series

2011. How Indefinites Choose Their Scope, coauthored with Donka Farkas ++ paper
A revised version of this ms. appeared in Linguistics and Philosophy.

The paper proposes a novel solution to the problem of scope posed by natural language indefinites that captures both the difference in scopal freedom between indefinites and bona fide quantifiers and the syntactic sensitivity that the scope of indefinites does nevertheless exhibit. Following the main insight of choice functional approaches, we connect the special scopal properties of indefinites to the fact that their semantics can be stated in terms of choosing a suitable witness. This is in contrast to bona fide quantifiers, the semantics of which crucially involves relations between sets of entities. We provide empirical arguments that this insight should not be captured by adding choice/Skolem functions to classical first-order logic, but in a semantics that follows Independence-Friendly Logic, in which scopal relations involving existentials are part of the recursive definition of truth and satisfaction. These scopal relations are resolved automatically as part of the interpretation of existentials. Additional support for this approach is provided by dependent indefinites, a cross-linguistically common class ofspecial indefinites that can be straightforwardly analyzed in our semantic framework.

2011. Sentence-Internal Different as Quantifier-Internal Anaphora ++ paper, related handout
A revised version of this ms. appeared in Linguistics and Philosophy.

The paper proposes the first unified account of deictic / sentence-external and sentence-internal readings of singular different. The empirical motivation for such an account is provided by a cross-linguistic survey and an analysis of the differences in distribution and interpretation between the English singular different, plural different and same (be it singular or plural). The main proposal is that distributive quantification temporarily makes available two discourse referents within its nuclear scope, the values of which are required by sentence-internal uses of singular different to be distinct, much as its deictic uses require the values of two discourse referents to be distinct. The contrast between singular different, plural different and same is explained in terms of several kinds of quantificational distributors that license their internal readings. The analysis is executed in a stack-based dynamic system couched in classical type logic, so we get compositionality in the usual Montagovian way; quantificational subordination and dependent indefinites in various languages provide additional motivation for the account. Investigating the connections between items with sentence-internal readings and the quantificational licensors of these readings opens up a larger project of formally investigating the typology of quantificational distributors and distributivity-dependent items and the fine-grained contexts of evaluation needed to capture this typological variation.

2011. Ranking and Necessity, coauthored with Alan Prince ++ paper (available as ROA 794)
A revised version of this ms. appeared in Natural Language & Linguistic Theory; related abstract, slides, handout; original 2003 ms.: Minimal Fusion Normal Form

Understanding a linguistic theory within OT requires an exact characterization of the ranking conditions necessitated by data. We introduce an algorithm (‘Fusional Reduction’, FRed) which calculates the necessary and sufficient ranking conditions inherent in any collection of candidates and presents them in a maximally concise and informative way. The algorithm, stemming from the original proposal of Brasoveanu (2003), develops in the setting of the fusional ERC theory of Prince (2002). In this context, the Most Informative Basis and the Skeletal Basis emerge as the two important types of reduced representations of ranking structure. We examine their properties and show how FRed produces them from data. Fine-grained FRed is compared with broad-stroke RCD (Tesar and Smolensky 1993, Tesar 1995 et seq.), and RCD is re-interpreted and embraced within FRed as a simplified, information-losing sub-case. Finally, FRed is compared with other related algorithms in structure, worst-case complexity, and relevance to the analytical enterprise.

2010. Dependent Indefinites within a Typology of Specificity, coauthored with Donka Farkas ++ handout
Presented at Specificity from Theoretical and Empirical Points of View, Stuttgart, Aug. 31-Sept. 2 2010.

2010. Plural Discourse Reference
A revised version of this ms. appeared in the Handbook of Logic and Language, 2nd edition, Johan van Benthem and Alice ter Meulen (eds.), Eline van der Ploeg and Jakub Szymanik (ed. assistants).

2010. Decomposing Modal Quantification ++ paper
A revised version of this ms. appeared in Journal of Semantics.

Providing a compositional interpretation procedure for discourses in which descriptions of complex dependencies between interrelated objects are incrementally built is a key challenge for natural language semantics. This paper focuses on the interactions between the entailment particle therefore, modalized conditionals and modal subordination and shows that the dependencies between individuals and possibilities that emerge out of such interactions can receive a unified compositional account in a system couched in classical type logic that integrates and simplifies van den Berg’s Dynamic Plural Logic and the classical Lewis-Kratzer analysis of modal quantification. The main proposal is that modal quantification is a composite notion, to be decomposed / analyzed in terms of discourse reference to quantificational dependencies that is multiply constrained by the various components that make up a modal quantifier. The system captures the truth-conditional and anaphoric components of modal quantification in an even-handed way and, unlike previous accounts, makes the propositional contents contributedby modal constructions available for subsequent discourse reference.

2010. Crossing the Appositive/At-issue Meaning Boundary, coauthored with Scott AnderBois and Robert Henderson ++ abstract, handout, paper
In the Proceedings of Semantics and Linguistic Theory 20 (UBC and SFU, Vancouver, April-May 2010).

This paper’s goal is to provide systematic evidence from anaphora, presupposition and ellipsis that appositive meaning, e.g. as contributed by the relative appositive in (1) below, and at-issue meaning, contributed by the main clause in (1), have to be integrated into a single, incrementally evolving semantic representation. While previous literature has provided partial arguments to this effect (Nouwen 2007 for anaphora, Amaral et al 2007 for both anaphora and presupposition), the systematic nature of this evidence — in particular, the evidence from ellipsis we will introduce — has been previously unnoticed.

1. John, who nearly killed a woman with his car, visited HER in the hospital.

We propose an analysis of these phenomena that integrates the dynamic account of anaphora and ellipsis as discourse reference to individuals and properties (respectively) with an account of at-issue meaning as a proposal to update the input Context Set (CS, see Stalnaker 1978)and of appositive meaning as an actual update of the CS that is not up for negotiation.

2010. Structured Anaphora to Quantifier Domains ++ paper
A revised version of this ms. appeared in Information and Computation.

The paper proposes a novel analysis of quantificational subordination, e.g. Harvey courts a woman at every convention. {She is very pretty. vs. She always comes to the banquet with him.} (Karttunen 1976), in particular of the fact that the indefinite in the initial sentence can have wide or narrow scope, but the first discourse as a whole allows only for the wide scope reading, while the second discourse allows for both readings. The cross-sentential interaction between scope and anaphora is captured in terms of structured anaphora to quantifier domains, formalized in a new dynamic system couched in classical type logic; given the underlying type logic, Montague-style compositionality at sub-clausal level follows automatically. Modal subordination (Roberts 1987) is analyzed in a parallel way, thereby capturing the parallels between the individual and modal domains argued for in Stone (1999). Several other phenomena are analyzed in terms of structured anaphora: exceptional wide scope, weak / strong donkey readings, anaphoric / uniqueness-implying definite descriptions and interactions between same / different and quantifier scope.

2010. Modal Concord as Modal Modification, coauthored with Pranav Anand ++ abstract, handout, paper
In the Proceedings of Sinn und Bedeutung 14 (also presented at IATL 25).

We argue that modal concord is best understood as modal modification by the adverb in question. Modal concord in deontic environments arises because the deontic adverb is a modifier that makes its own ancillary modal claim. Modal flavor consonance is analyzed as a grammatically determined phenomenon, while agreement in modal force is taken to be a pragmatically mediated phenomenon. The paper concludes with the outline of a broader (corpus-driven) investigation of the lexical semantics of modal adverbs and their meaning-based classification and distributionand argues that the investigation of the standard cases of modal concord should be analyzed as part of this broader project.

2010. Modified Numerals as Post-suppositions ++ abstract, a detailed handout, paper
Presented at Amsterdam Colloquium 17, Goethe-Universität Frankfurt am Main and SCALE: A workshop on Scales, Comparison and their Linguistic Expression (Stanford).

The goal of this paper is to provide a compositional account of cumulative readings with non-increasing modified numerals (aka van Benthem’s puzzle) — e.g., Exactly three boys saw exactly five movies — in terms of cardinality post-suppositions. The reading of this sentence that we want to capture is the cumulative one, namely: consider the maximal number of boys that saw a movie and the maximal number of movies seen by a boy; there are three such boys and five such movies. Importantly, this reading is different from: the maximal number of boys that saw exactly five movies is three. The main proposal is that modified numerals make two kinds of contributions to the meaning of sentences like the one above. Their asserted / at-issue contribution is a maximization operator that introduces the maximal set of entities that satisfies their restrictor and nuclear scope. The second contribution is a post-supposition, i.e., a cardinality constraint (e.g., exactly three) that needs to be satisfied relative to the context that results after the at-issue meaning is evaluated. For our current purposes, contexts are sets of variable assignments relative to which quantificational expressions are interpreted — and which are updated as a result of the interpretation of such expressions. The main difference between the present account and Krifka (1999) is conceptual: we take modified numerals to constrain quantificational — and not focus — alternatives, where a quantificational alternative is one of the sets of assignments satisfying a quantificational expression.

2009. Varieties of Distributivity: One by One vs Each, coauthored with Robert Henderson ++ abstract, handout, paper
Presented at Semantics and Linguistic Theory 19 (OSU, April 2009) and TRilateral linguistics weekEND (TREND; Stanford, May 2009).

The main goal of the paper is to argue that distributive quantificational dependencies in natural language can be established in two different ways: (i) by encapsulating quantification into functions storing quantificational dependencies as a whole, needed to account for one by one-based distributive sentences like The boys recited ‘The Raven’ one by one — and (ii) by decomposing quantification in such a way that each n-tuple of quantificationally dependent entities is individually stored in a variable assignment and quantifiers are interpreted relative to the entire set of variables assignments that stores quantificational dependencies in this pointwise / assignmentwise manner — needed to account for each-based distributive sentences like The boys each recited a different poem.

2009. Scope and the Grammar of Choice, coauthored with Donka Farkas ++ handout, paper
In the Proceedings of the Tenth Symposium on Logic and Language (LoLa 10); also presented at California Universities Semantics & Pragmatics (CUSP; UCLA, May 2009) and the Workshop on Language, Communication and Rational Agency (Stanford, May 2009).

The paper proposes a novel solution to the problem of scope posed by natural language indefinites that captures both the difference in scopal freedom between indefinites and quantifiers — indefinites have free upwards scope, disregarding not only clausal but also island boundaries — and the fact that the scopal freedom of indefinites is nonetheless syntactically constrained. Following the main insight of independence-friendly logic, the special scopal properties of indefinites are attributed to the fact that their semantics can be stated in terms of choosing a suitable witness. This is in contrast to “bona fide” quantifiers, the semantics of which crucially involves relations between sets of entities. The syntactic constraints on the interpretation of indefinites follow from the fact that witness choice arises as a natural consequence of the process of (syntax-based) compositional interpretation of sentences and it is not encapsulated into the lexical meaning of indefinites, as choice / Skolem function approaches to exceptional scope would have it. An analysis of dependent indefinites is provided using the same framework and formal ingredients.

2008. Donkey Pluralities: Plural Information States vs. Non-atomic Individuals ++ paper
A revised version of this ms. appeared in Linguistics and Philosophy.

The paper argues that two distinct and independent notions of plurality are involved in natural language anaphora and quantification: plural reference (the usual non-atomic individuals) and plural discourse reference, i.e., reference to a quantificational dependency between sets of objects (e.g., atomic/non-atomic individuals) that is established and subsequently elaborated upon in discourse. Following van den Berg (1996), plural discourse reference is modeled as plural information states (i.e., as sets of variable assignments) in a new dynamic system couched in classical type logic that extends Compositional DRT (Muskens 1996). Given the underlying type logic, compositionality at sub-clausal level follows automatically and standard techniques from Montague semantics become available. The idea that plural info states are semantically necessary (in addition to non-atomic individuals) is motivated by relative-clause donkey sentences with multiple instances of singular donkey anaphora that have mixed (weak and strong) readings. At the same time, allowing for non-atomic individuals in addition to plural info states enables us to capture the intuitive parallels between singular and plural (donkey) anaphora, while deriving the incompatibility between singular (donkey) anaphora and collective predicates. The system also accounts for empirically unrelated phenomena, e.g., the uniqueness effects associated with singular (donkey) anaphora discussed in Kadmon (1990) and Heim (1990) among others.

Here are several kinds of (relative-clause) donkey sentences that are analyzed in the paper:

1. Every person who buys a book on amazon.com and has acredit card uses it to pay for it. (mixed weak & strong donkey sentences)
2. # Every farmer who owns adonkey gathers it around the fire at night. (singular donkey anaphora and collective predicates)
3. Every parent who gives a balloon to two boys expects them to end up fighting (each other) for it. (multiple singular and plural donkey anaphora and collective predicates)
4. Everybody who bought two sage plants here bought seven others along with them. (plural sage plant examples)
5. Most house-elves who fall in love with a witch buy her an alligator purse. (proportions)
6. Every man who has a son wills him all his money. (uniqueness effects and donkey anaphora)
7. No man who had acredit card failed to use it. (strong donkey readings with no)
8. Every person who had a dime in his pocket refused to put it in the meter. (strong readings for dime examples with nuclear scope negation)
9. Every company that hired a Moldavian man, but no company that hired a Transylvanian man promoted him within two weeks of hiring. (mixed weak & strong donkey sentences with a single donkey pronoun)
10. Every man who introduced a friend to me thought we had something in common. (plural donkey anaphora with split antecedents)
11. Every linguist who works on a difficult problem is interested to read most papers that were written about it. (donkey anaphora and exceptional wide scope)

2008. Comparative Correlatives as Anaphora to Differentials ++ abstract, poster, slides, handout, paper
In the Proceedings of Semantics and Linguistic Theory 18 (UMass Amherst, 2008).

The empirical goal of the paper is to establish that there are comparative correlatives that are not comparative conditionals, against what much of the previous literature assumes. This is shown by the Romanian comparative correlative in (1) and is further supported by the equative correlative in (2). No conditional paraphrase is possible for (1) or (2) - as, for example, Beck (1997) (“On the Semantics of Comparative Conditionals”) would have it - since they are statements about what is actually the case. In contrast, the comparative correlative in (3) can be paraphrased by a conditional, e.g. (on one of its readings): if a lawyer x is more aggressive than a lawyer y by a certain amount, then x is more efficient than y by a corresponding amount. The main proposal is that a unified analysis should be given for such non- conditional, differential-based comparative (and equative) correlatives and the more familiar, conditional-like comparative correlatives in terms of a relation (possibly the identity relation) between differentials. In paraticular, the Romanian atît (that much) is anaphoric to differential intervals, i.e., atît is a proform in the degree domain - and the wh- differential cît (how much) is an indefinite introducing a non-empty interval, anaphorically retrieved by atît.

1. Cu cît e mai înalt frate-le decît sora, cu atît e mai înalt tată-l decît mama. With how much is more tall brother-the than sister.the, with that much is more tall father-the than mother.the The brother is taller than the sister by a certain amount and the father is taller than the mother by the same amount.
2. Pe cît e Irina de frumoasă, pe atît e de deşteaptă. PE how much is Irina DE beautiful, PE that much is DE smart Irina is beautiful to a certain, significant extent and she is smart to the same, equally significant extent.
3. Cu cît e un avocat mai agresiv, cu atît e mai eficient. With how much is a lawyer more aggressive, with that much is more efficient The more aggressive a lawyer is, the more efficient s/he is.

2008. Measure Noun Polysemy and Monotonicity: Evidence from Romanian Pseudopartitives ++ paper, abstract, extended abstract, handout, poster
In the Proceedings of the 38th Meeting of the North East Linguistic Society (also presented at the Syntax and Semantics of Measurability Work shop, CASTL).

The two goals of this paper are: (i) to argue that, syntactically, the measure noun is the head of the extended projection in Romanian pseudopartitive constructions like (1) - much like the leftmost noun is the head of true partitive constructions like (2) (in Romanian, the preposition de appears only with pseudopartitives, while the preposition din/dintre appears only with true partitives); (ii) to propose a suitable semantics for pseudopartitives that accommodates this syntactic generalization.

1. zece grame de brînză (de capră) ten grams of cheese (of goat) ten grams of (goat) cheese
2. zece grame din această brînză (de capră) ten grams of this cheese (of goat) ten grams of this (goat) cheese
3. # zece grame din brînză (de capră)
4. # zece grame de această brînză (de capră)

One of the two main contributions is arguing that measure nouns are polysemous, i.e. they have two distinct, but closely related senses: (i) a degree-based one, present in comparatives like Linus is two pounds heavier than Gabby or (arguably) nominal compounds like two pound stone, and (ii) an individual-based sense, present in pseudopartitives like ten grams of cheese or true partitives like ten grams of this cheese and their Romanian counterparts in (1) and (2) above. Secondly, the polysemy proposal enables us to derive the observation in Schwarzschild (2006) (“The Role of Dimensions in the Syntax of Noun Phrases”, Syntax 9.1) that measure expressions are monotonic in pseudopartitives (I use (non-)monotonic in the sense of Schwarzschild 2006). Syntactically and semantically, the measure expression is the head of the pseudopartitive while the other nominal expression is the non-head, in contrast to Schwarzschild (2006), where the head/non-head categorization is reversed.

2007. Say Reports, Assertion Events and Meaning Dimensions, coauthored with Donka Farkas ++ paper
In Pitar Moş: A Building with a View. Papers in Honour of Alexandra Cornilescu, G. Alboiu, A. Avram, L. Avram & D. Isac (eds.), Bucharest: Editura Universităţii din Bucureşti.

In this paper, we study the parameters that come into play when assessing the truth conditions of say reports and contrast them with belief attributions. We argue that these conditions are sensitive in intricate ways to the connection between the interpretation of the complement of say and the properties of the reported speech act. There are three general areas this exercise is relevant to, besides the immediate issue of understanding the meaning of say: (i) the discussion shows the need to go beyond the simplest view of propositional attitudes, which treats them as restricted quantifiers over worlds; (ii) the complex connections that must exist between the say report and its source speech act show that one has to be able to differentiate between various layers of meaning for the antecedent sentences; (iii) finally, this paper is a small step towards a typology of propositional attitudes that allows us to uncover the complex web of relationships that grammatical mood is sensitive to.

2006 (December). Structured Nominal and Modal Reference (1up pdf, 378pp) ++ PhD dissertation, Rutgers University. Committee: Maria Bittner (chair), Hans Kamp (external member), Roger Schwarzschild and Matthew Stone.

The dissertation argues that discourse reference in natural language involves two equally important components with essentially the same interpretive dynamics, namely reference to values, i.e. non-singleton sets of objects (individuals and possible worlds), and reference to structure, i.e. the correlation / dependency between such sets, which is introduced and incrementally elaborated upon in discourse.

To define and investigate structured discourse reference, a new dynamic system couched in classical (many-sorted) type logic is introduced which extends Compositional DRT (CDRT, Muskens 1996) with plural information states, i.e. information states are modeled as sets of variable assignments (following van den Berg 1996a), which can be represented as matrices with assignments (sequences) as rows. A plural info state encodes both values (the columns of the matrix store sets of objects) and structure (each row of the matrix encodes a correlation / dependency between the objects stored in it). Given the underlying type logic, compositionality at sub-clausal level follows automatically and standard techniques from Montague semantics (e.g. type shifting) become available.

The idea that plural info states are semantically necessary is motivated by examples with morphologically singular anaphors, in contrast to the previous literature that argues for plural info states based on plural anaphora. Plural Compositional DRT (PCDRT) enables us compositionally account for a variety of phenomena, including: (i) mixed weak & strong donkey anaphora, e.g., Every person who buys a computer and has a credit card uses it to pay for it, (ii) quantificational subordination, e.g., Harvey courts a woman at every convention. She always comes to the banquet with him (Karttunen 1976), (iii) modal anaphora and modal subordination, e.g., A wolf might come in. It would eat Harvey first (based on Roberts 1989) and (iv) naturally-occurring discourses exhibiting complex interactions between modalized conditionals, donkey anaphora, modal subordination and the entailment particle therefore, e.g., [A] man cannot live without joy. Therefore, when he is deprived of true spiritual joys, it is necessary that he become addicted to carnal pleasures (Thomas Aquinas).

The PCDRT account of these phenomena explicitly and systematically captures the anaphoric and quantificational parallels between the individual and modal domains.

Table of Contents

• Abstract, Acknowledgments, Table of Contents
• Chapter 1. Introduction
• Chapter 2. Dynamic Predicate Logic with Generalized Quantification
  • Dynamic Predicate Logic (DPL)
  • Anaphora in DPL
  • Extending DPL with Unselective Generalized Quantification
  • Extending DPL with Selective Generalized Quantification (DPL+GQ)
  • Limitations of DPL+GQ: Mixed Weak & Strong Donkey Sentences
• Chapter 3. Compositional DRT
  • Dynamic Ty2
  • Translating DPL into Dynamic Ty2
  • Syntax of a Fragment of English
  • Type-driven Translation
  • Anaphora and Quantification in Compositional DRT (CDRT)
• Chapter 4. Compositional DRT with Generalized Quantification
  • Translating Unselective Quantification into Dynamic Ty2
  • Translating Selective Quantification into Dynamic Ty2
  • Extending CDRT with Generalized Quantification (CDRT+GQ)
  • Anaphora and Generalized Coordination in CDRT+GQ
  • Limitations of CDRT+GQ: Mixed Weak & Strong Donkey Sentences
• Chapter 5. Structured Nominal Reference: Donkey Anaphora
  • Outline of the Proposed Account
  • CDRT+GQ with Plural Information States: Plural CDRT (PCDRT)
  • Solutions to Donkey Problems
  • Comparison with Alternative Approaches
• Chapter 6. Structured Nominal Reference: Quantificational Subordination
  • Structured Anaphora to Quantifier Domains
  • Redefining Generalized Quantification
  • Quantificational Subordination in PCDRT
  • Comparison with Alternative Approaches
• Chapter 7. Structured Modal Reference: Modal Anaphora and Subordination
  • Structured Reference across Domains
  • Intensional Plural CDRT (IP-CDRT)
  • Conditionals, Modals and Therefore in IP-CDRT
  • Modal Subordination in IP-CDRT
  • A Parallel Account of Modal and Quantificational Subordination
  • Comparison with Alternative Approaches
• Chapter 8. Conclusion
• References

2006. Temporal and Propositional De Se: Evidence from Romanian Subjunctive Mood ++ slides, paper, an older abstract
In the Proceedings of Sinn und Bedeutung 10, published as ZASPIL (ZAS Papers in Linguistics), Volume 44, Ch. Ebert & C. Endriss (eds.), 2006.

The paper investigates the interpretation of the Romanian subjunctive B (subjB) mood when it is embedded under the propositional attitude verb crede (believe). SubjB is analyzed as a single package of three distinct presuppositions: temporal de se, dissociation and propositional de se. I show that subjB is the temporal analogue of null PRO in the individual domain: it allows only for a de se reading. Dissociation enables us to show that subjB always takes scope over a negation embedded in a belief report. Propositional de se derives this empirical generalization. The introduction of centered propositions (generalizing centered worlds), together with propositional de se, dissociation and the belief ‘introspection’ principles, derives the fact that subjB belief reports (unlike their indicative counterparts) are infelicitous with embedded probabil.

2005. Contrast Analysis Aids The Learning of Phonological Underlying Forms, coauthored with John Alderete, Nazarré Merchant, Alan Prince and Bruce Tesar ++ abstract, paper
In the Proceedings of the 24th West Coast Conference on Formal Linguistics (WCCFL 24), J. Alderete et al (eds.), Cascadilla Proceedings Project, Cascadilla Press.

One of the many challenges to be faced in explaining language learning is the interdependence of the phonological mapping and the phonological underlying forms for morphemes. The learner must attempt to infer both simultaneously, based on the surface forms of a language. The present paper presents evidence in support of the idea that observations about surface contrast can inform the learner about the content of underlying forms. More specifically, contrasting outputs for morphemes in a given environment can provide information about underlying forms. We present one way of capitalizing on such information in a learning algorithm, and show how contrast information can combine with phonotactic information to aid learning.