Archive
Spring 2024
Feb 6
Diamonds are a disjunction's best friend
I introduce a number of new puzzles about the relation between disjunction and possibility. I argue that these puzzles are best solved with a theory on which 'p or q' means (p∨q)∧<>p∧<>q, where <> is a possibility modal underspecified for modal flavor. I show that the resulting theory also yields an elegant account of free choice inferences.
Feb 29
Entity Tracking in Language Models
Keeping track of how states of entities change as a text or dialog unfolds is a key prerequisite to discourse understanding. We propose a behavioral task testing to what extent a language model can infer the final state of an entity given a natural language description of the initial state and a series of state-changing operations, following a set of desiderata we lay out for measuring nontrivial entity tracking capacity. Our evaluations of several language models reveal that only (1) in-context learning with models trained on large amounts of code, or (2) finetuning a model directly on the entity tracking task lead to nontrivial entity tracking behavior. This suggests that language models can learn to track entities but pretraining on text corpora alone does not make this capacity surface. In light of these results, I will end with brief discussions of ongoing work investigating the role of code training further, as well as testing for latent representations of entity states.
April 3
The conceptual structure of the word belief
Within a mentalist approach to semantics, the meanings of words and sentences pertain not to the real world (or a set of possible worlds), but to the world as conceptualized by a language user. The concept expressed by the word ‘belief’ is part of “folk psychology” or Theory of Mind – the way humans ascribe mental states to others. The issue for this talk is therefore where the concept of belief fits into the ecology of the Theory of Mind.
Using grammatical patterns not usually cited in the literature of philosophy of language, I will show that beliefs pattern grammatically and semantically very much like the description of pictures. In particular, the classical paradoxes pertaining to belief in the philosophical literature find exact parallels in the description of pictures. I conclude that beliefs are conceptualized as a sort of picture in the head, a representation of a state of affairs.
In addition, there are strong grammatical and semantic parallels between predicates pertaining to beliefs and predicates pertaining to intention. Again in the domain of intention, it is possible to reconstruct the standard paradoxes of belief. These parallels suggest that believing and intending are two sides of the same coin. Ascribing a belief to someone amounts to attributing to them a commitment to a state of affairs, while ascribing an intention to someone attributes to them a commitment to perform an action.
In short, the concept of ‘belief’ and its relatives is rich and ramified, and this structure can be discovered in part through detailed linguistic analysis.
April 9
Decision and Tenable Conditionals
I present a hybrid decision theory, coinciding sometimes with (traditional) EDT, but usually with (traditional) CDT, which is inspired by recent work on unified and fully compositional approaches to the probabilities of conditionals (Bacon, 2015; Goldstein & Santorio, 2021; Schultheis, forthcoming a). The hybrid theory features a few other loci of interest: the partitionality of options fails in an important way, and close attention is paid to how one might (dis)confirm chance hypotheses under the umbrella of the Principal Principle. On this theory, the probabilities of conditionals play a role in underwriting a theory of credal chance that follows Skyrms’s Thesis (Skyrms, 1981, 1984) about the probabilities of counterfactuals. Moreover, the credences it is epistemically rational to assign to these conditionals can guide updating on one’s own acts. This implies some departures from Conditionalization—departures I defend on epistemological grounds. This has important ramifications for cases of diachronic instability.
April 25
Controlled Rearing of Language Models can reveal Linguistic Insight!
Neural network based language models have overwhelmingly proliferated across a myriad of disciplines. Meanwhile, their precise role in advancing our knowledge about human cognition remains a topic of much debate. What can these black-boxes tell us about language learning, use, and generalization? In this talk I will discuss some approaches my colleagues and I have undertaken to contribute to this debate. In particular, I will discuss the paradigm of “controlled rearing” – the informed manipulation of the developmental environment of learners–in this case a language model. I will then describe how different ways of performing controlled rearing can be applied to test and generate hypotheses about the conditions necessary for generalization in domains covering both rare as well as well-known linguistic phenomena.
Fall 2023
Sept 22
Moorean Promises
"I promise to mow your lawn, but I don’t know whether I will." "I’ll go to your talk–that’s a promise! Of course, I might not make it." These "Moorean" promises sound absurd. But why? This talk proposes an answer. In the literature on assertion, many have used Moore’s paradox (e.g., "It's raining, but I don't whether it is") to motivate a strong epistemic condition on assertion, according to which one should only assert p if one knows p (or is epistemically certain of p). I put forward an analogous epistemic norm on promising, according to which one should only promise to do something if one knows (or is epistemically certain) one will do it. This norm explains why Moorean promises are absurd, and it makes sense of a range of further linguistic data. It also sheds light on the relationships between different speech acts, and helps explain why promises generate obligations.
Oct 6
Causation, implicativity, and the logic of ability
Against linguistic analyses which treat them as circumstantial possibilities, ability modals fail to verify certain axioms of alethic modality (T, K). The philosophical literature includes a number of 'conditional' analyses which aim to explain this behaviour by representing ability as a necessity modal conditioned on some potential action by the sentential subject (or agent of ability). Generally speaking, these approaches leave open the nature of the connection between the conditioning action and the ability target.
I suggest that the action-ability connection is best analyzed in terms of causal dependence, arguing that a complex causal structure for ability explains not only the perceived strength of ability (as compared to circumstantial possibility), but also offers an account of a longstanding crosslinguistic puzzle in the interpretation of ability under aspectual modification. As shown for French in (1)-(2), imperfectively-marked ability modals license a 'pure', potentially-unrealized ability interpretation, but their perfective counterparts give rise to actuality entailments, forcing the realization of the modal complement (Bhatt 1999).
1.Marja pouvait traverser lac à la nage, mais elle ne l'a jamais traversé.
'Marja could-IMPF swim across the lake, but she never crossed it.'
2. Marja a pu traverser le lac à la nage, #mais elle ne l'a pas traversé.
'Marja could-PFV swim across the lake, #but she did not cross it.’
I develop the causal account of ability by comparing actualized ability readings to the interpretation of two types of implicative predicates (lexical implicatives and enough comparatives), which license variable-strength inferences to the actualization of their complements. Building on prior work (Nadathur 2019, 2023), I show that implicative complement inferences follow from causal dependence relations embedded in the lexical and compositional structure of the predicate, and argue that the aspectual class properties of these causal structures interact with viewpoint aspect to produce contrasts paralleling (1)-(2). Extending the implicative treatment allows the interpretations in (1)-(2) to be derived from a single underlying analysis of ability, and suggests an important role for causal information in the semantic analysis of language that is not overtly causal in nature.
Nov 10
Representing Complex Situations
We use natural language to convey information about situations: things that happen or stuff that is true. This ability is supported by systematic relationships between the way we conceptualize situations and the way we describe them. These systematic relationships in turn underwrite inferences that go beyond what one strictly says in describing a situation. The question that motivates this talk is how to design systems that correctly capture the inferences we draw about situations on the basis of their descriptions.
Classical approaches to this question–exemplified in their modern form by graph-based representations, such Abstract Meaning Representation–attempt to capture the situation conceptualization associated with a description using a symbolic situation ontology and to draw inferences on the basis of rules stated over that ontology. An increasingly popular alternative to such ontology-factored approaches are ontology-free approaches, which attempt to directly represent inferences about a situation as natural language strings associated with a situation description, thereby bypassing the problem of engineering a situation ontology entirely.
I discuss the benefits and drawbacks of these two approaches and present two case studies in synthesizing them that focus specifically on how best to capture inferences about complex situations–i.e. situations, like building a house, that themselves may be composed of subsituations, like laying the house’s foundations, framing the house, etc. I argue that we should ultimately strive for ontology-free representations but that the challenges inherent to reasoning about complex situations highlight the persistent benefits of situation ontologies in providing representational scaffolding for the construction of such representations.
Nov 17
A Computational Approach to Questions Under Discussion in Discourse Modeling
Discourse structures characterize inter-sentential relationships and textual organization, enhancing high-level text comprehension. However, deriving these structures relies on annotated data that are linguistically sophisticated and thus challenging to obtain. In the wake of large language models, we introduce a paradigm shift that views discourse structure through the lens of free-form question answering, aligning with the linguistic framework of Questions Under Discussion (QUD). We construct two tasks under QUD: (1) generating curiosity-driven, open-ended questions and locating their answers in a document; (2) QUD dependency parsing, by considering each sentence as the answer to an implicit question elicited from context. We further discuss the evaluation of QUD question generation in a linguistically informed scheme. We conclude with potential new applications enabled by the QUD framework that highlight its versatility, and natural fit with LLMs and human interactions alike.
Dec 1
Extensionality and Quantification
One might have thought that, once one has adopted the standard definition of a well-formed formula and Tarski's recursive definition of satisfaction, the question whether the quantifiers create extensional or non-extensional contexts would have a definite, indisputable answer. Yet in the philosophical and logical literature we find apparently well-reasoned but nevertheless conflicting answers to this question. In the talk, I l will elucidate why this is and, after a detour through the logical analysis of a natural-language phenomenon, try to assess the merits of extensionalism and non-extensionalism with respect to quantification.
Spring 2023
February 9
Neo-Pragmatism about Truth
Deflationists about truth hold that the function of the truth predicate is to enable us to make certain assertions we could not otherwise make. Pragmatists claim that the utility of negation lies in its role in registering incompatibility. The pragmatist insight about negation has been successfully incorporated into bilateral theories of content, which take the meaning of negation to be inferentially explained in terms of the speech act of rejection. One can implement the deflationist insight in the pragmatist’s theory of content by taking the meaning of the truth predicate to be explained by its inferential relation to assertion. There are two upshots. First, a new diagnosis of the Liar, Revenges and attendant paradoxes: the paradoxes require that truth rules preserve evidence, but they only preserve commitment. Second, one straightforwardly obtains axiomatisations of several supervaluational hierarchies, answering the question of how such theories are to be naturally axiomatised. This is joint work with Luca Incurvati (Amsterdam).
February 17
Why Communicative Intentions?
Grice argued that we communicate by revealing our intentions to change our addressees’ states of mind. I agree, but I don’t think that Grice did a good job of justifying this view or explaining why we would communicate in this way. The why question is important because if Grice is right, ordinary communication is cognitively costly—a point that has stoked criticisms of Grice and inspired alternative models. Psycholinguistic evidence suggests that ordinary human communication does involve lots of cognitively costly reasoning. We pay these costs, I will argue, because it’s worth it for the massive communicative benefits that we get in return. My explanation of these benefits has two parts, corresponding to the first two components of a communicative intention. First: We form intentions to change others’ states of mind as a crucial step in the practical reasoning by means of which we design communicative acts for our addressees. This design process greatly increases the flexibility and reliability of human communication, and allows us to speak natural languages with far greater expressive power. Second: We reveal these intentions to our addressees because this is an important precondition for treating communication as a cooperative joint activity. When this works, it means that a communicator can rely on their addressee to invest their own cognitive resources to the project of understanding and assessing what they say, thereby dramatically expanding the sphere of communicative goals that it is possible to achieve.
March 2
Coordinated on the context: Discourse salience, exclusivity, mirativity, precisification, and intensification in Marathi (Indo-Aryan)
According to Lewis (1969), coordination problems involve situations where agents must decide upon the same strategy from among a set of alternatives in order to achieve desirable outcomes. Schelling (1960) showed that in many such cases, even in novel coordination problems, where agents have no prior knowledge of each other’s preference or experience, they tend to zero-in on one strategy that seems most “natural” or “salient”. These are focal points or Schelling points. Such successful coordination implies that at least for some coordination problems, agents’ expectations about one another’s behavior are crucial in determining the optimal solution.
In this talk, I discuss a chameleon-like discourse clitic that is ubiquitous in Indo-Aryan languages (occurring in at least Bangla, Gujarati, Hindi, and Marathi). The uses of this clitic overlap with those of English particles like exclusives only/just, intensifiers really/totally, precisifiers right/exactly/absolutely, anaphoric indeed/that very, and scalar additive even without corresponding perfectly to any of them. I offer an analysis of the varied and seemingly disparate uses of this expression, focusing on the Marathi variant =ts. My claim is that =ts conventionally signals that the proposition expressed by the prejacent is uniquely salient to the interlocutors among alternatives in the Current Question (CQ). In other words, =ts explicitly conveys that the proposition expressed by the prejacent offers a Schelling point for the interlocutors to coordinate on. Most discourse effects associated with =ts are shown to arise as a consequence of pragmatic reasoning about the position of the prejacent with respect to the contextually given ordering on the CQ. In addition to offering a unified analysis for Marathi =ts and its functional cognates in Indo-Aryan, this new perspective can open the door to a better understanding of why meanings like exclusivity, mirativity, precisification, and intensification might cluster together in the same lexical item across languages.
April 6
Free choice and presuppositional exhaustification
Sentences such as Olivia can take Logic or Algebra (‘♦∨-sentences’) are typically assigned the ‘Free Choice’ (FC) reading that Olivia can take Logic and can take Algebra. Given a standard semantics for modals and disjunction, such FC readings are not predicted from the surface form of ♦∨-sentences. An influential approach treats FC readings as a kind of scalar enrichment generated by a covert exhaustification operator. This approach can also account for the ‘double prohibition’ readings of ¬♦∨-sentences like Olivia can’t take Logic or Algebra via general principles of implicature cancellation in downward entailing environments. Marty & Romoli (2020) and Romoli & Santorio (2019) examine the projection and filtering behavior of embedded ♦∨ and ¬♦∨-sentences, focusing on two kinds of cases which challenge this influential approach. First, ♦∨-sentences under negative factives: e.g., Noah is unaware that Olivia can take Logic or Algebra, which in its most salient reading presupposes that Olivia has free choice and yet attributes to Noah ignorance not just concerning whether Olivia has free choice but also whether she can take even one of the classes. Second, ¬♦∨-sentences embedded under disjunction: e.g., Either Maria can’t study in Tokyo or Boston, or she is the first in our family who can study in Japan and the second who can study in the States, which in its most salient reading filters out the FC presupposition of the second disjunct while the first disjunct gets the standard double prohibition reading. These sentences present a serious challenge to extant accounts of FC. In this talk, we present a novel exhaustification-based account of FC that issues in a uniform solution to Marty, Romoli and Santorio’s puzzles concerning the presuppositional and filtering behavior of embedded ♦∨, ¬♦∨, and related FC sentences. Our account builds on the proposal—advanced in Bassi et al. (2021) and Del Pinal (2021) as a general theory of scalar implicatures—that covert exhaustification is a presupposition trigger such that the prejacent forms the assertive content while any excludable (or includable) alternatives are incorporated at the non-at issue, presuppositional level.
April 20
Syntax and semantics in the age of large language models
The recent rise in large language models profoundly changes the landscape for theories of human language. I'll discuss how these models should cause us to rethink many popular ideas about grammar, including most prominently those argued for by Chomsky. I'll also discuss the way in which these models implement theories of language and grammar, as well as the links and gaps between these models and child language learning. Despite important differences, I'll argue that people who care about learning should take LLMs seriously. Finally, these models provide a compelling way to think about semantics and conceptual representation. I'll argue that the sense in which they possess "meanings" is likely to be analogous to how human words and concepts achieve meaning.
Fall 2022
September 22
Unnatural Language Semantics
Unnatural language semantics is the study of the meaning of words and expressions in languages that are very unlike natural languages. In this talk, I will present several case studies about how unnatural language semantics can inform us about the structure of natural languages. In particular, I will explain and present several case studies of two methods of explaining semantic universals (shared properties of meaning across the languages of the world): one arguing that such universals arise due to learnability, and another due to optimally trading-off the competing pressures of simplicity and informativeness. The talk will conclude with some discussion about the relative merits of the two explanations as well as other avenues where unnatural language semantics can be helpful.
September 29
Competing influences on word meaning in context
In this talk, I argue for a notion of word meaning in context that characterizes meaning as both intensional and conceptual. I introduce a framework for specifying local as well as global influences on word meaning in context, together with their interactions. In this framework, sentence meaning is represented through a Situation Description System, a probabilistic model which takes utterance understanding to be the mental process of describing to oneself one or more situations that would account for an observed utterance. I will show analyses of sentences containing various contextualization phenomena.
October 6
Jonathan Caleb Kendrick
Question Sensitive Abilities
There’s an asymmetry in strength between ability can and its negation can’t. While it’s perfectly acceptable to say “I can φ, but I won’t φ,” it’s unacceptable to assert “I can’t φ, but I will φ.” This suggests that, if you can’t φ, then you won’t φ–in other words, can’t φ entails ¬φ. However, this principle appears incompatible with another seemingly true principle. As Kenny (1975) observed, φing doesn’t always entail an ability to φ–hitting a bullseye by sheer luck doesn’t entail an ability to hit bullseyes. But, φ entails can φ is simply the contrapositive of can’t φ entails ¬φ. We resolve this apparent tension by developing a trivalent semantics where ability can is treated as a quantifier over answers to a salient deliberative question.
October 27
“Understanding” and prediction: Disentangling meaning extraction and predictive processes in humans and AI
The interaction between "understanding" and prediction is a central theme both in psycholinguistics and in the AI domain of natural language processing (NLP). Evidence indicates that the human brain engages in predictive processing while extracting the meaning of language in real time, while NLP models use training based on prediction in context to learn strategies of language "understanding". In this talk I will discuss work that tackles key problems in both of these domains by exploring and teasing apart effects of compositional meaning extraction and effects of statistical-associative processes associated with prediction. I will begin with work that diagnoses the linguistic capabilities of NLP models, investigating the extent to which these models exhibit robust compositional meaning processing resembling that of humans, versus shallower heuristic sensitivities associated with predictive processes. I will show that with properly controlled tests, we identify important limitations in the capacities of current NLP models to handle compositional meaning as humans do. However, the models' behaviors do show signs of aligning with statistical sensitivities associated with predictive mechanisms in human real-time processing. Leveraging this knowledge, I will then turn to work that directly models the mechanisms underlying human real-time language comprehension, with a focus on understanding how the robust compositional meaning extraction processes exhibited by humans interact with probabilistic predictive mechanisms. I will show that by combining psycholinguistic theory with targeted use of measures from NLP models, we can strengthen the explanatory power of psycholinguistic models and achieve nuanced accounts of interacting factors underlying a wide range of observed effects in human language comprehension.
November 10
Evidence and Conditional Propositions
This paper is about the epistemological upshots of Stalnaker’s Thesis, the thesis that, very roughly, the probability in a conditional proposition A >C given some body of evidence should be equal to the probability of C, conditional on A, given that evidence. A leading way to deliver Stalnaker’s Thesis assumes a pluralist view of conditional propositions, where for each possible evidential accessibility relation there is a corresponding conditional proposition. I show that, given plausible background assumptions, conditional pluralism and its version of Stalnaker’s Thesis entails the Negation Introspection principle for evidence. I reject this conclusion and so reject the pluralist view of conditionals. In its place I develop a monist information-sensitive view of conditionals. The view is monist because there is just one conditional which yields Stalnaker’s Thesis; the view is information-sensitive, because what conditionals are part of our evidence is fully determined by what factual propositions are part of our evidence. I show that this view is able to validate Stalnaker’s Thesis without the introspection principles.