%{search_type} search results

• 1. Introduction
• 2. A Preliminary Case for Conventional Implicatures
• 3. A Logic for Conventional Implicatures
• 4. Supplements
• 5. Expressive Content
• 6. The Supplement Relation: A Syntactic Analysis
• 7. A Look Outside Grice's Definition
• Appendix
• Bibliography.
• (source: Nielsen Book Data)

This book revives the study of conventional implicatures in natural language semantics. H. Paul Grice first defined the concept. Since then his definition has seen much use and many redefinitions, but it has never enjoyed a stable place in linguistic theory. Christopher Potts returns to the original and uses it as a key into two presently under-studied areas of natural language: supplements (appositives, parentheticals) and expressives (e.g., honorifics, epithets). The account of both depends on a theory in which sentence meanings can be multidimensional. The theory is logically and intuitively compositional, and it minimally extends a familiar kind of intensional logic, thereby providing an adaptable, highly useful tool for semantic analysis. The result is a linguistic theory that is accessible not only to linguists of all stripes, but also philosophers of language, logicians, and computer scientists who have linguistic applications in mind.
(source: Nielsen Book Data)

Speakers exhibit considerable production variability at all levels of linguistic representations. This raises the question how successful communication is nevertheless possible most of the time. In this dissertation, I investigate this question and I study to what extent listeners adapt to variable use of words using the example of uncertainty expressions such as 'might' and 'probably'. In several web-based experiments, I show that listeners exhibit uncertainty in their expectations about a generic speaker's use of uncertainty expressions; that listeners update production expectations to match a specific speaker's use of uncertainty expressions after a brief exposure to that speaker; and that updated production expectations result in updated speaker-specific interpretations of uncertainty expressions. I further investigate the associated cognitive processes and I investigate what kind of representations listeners update during semantic-pragmatic adaptation. To this end, I present a novel Bayesian computational model of production expectations of uncertainty expressions and a novel model of the adaptation process based on Bayesian belief updating. Through a series of simulations, I find that post-adaptation behavior is best predicted by a model that assumes that listeners update both speaker-specific semantic representations and speaker-specific utterance choice preferences, suggesting that listeners update at least these two types of representations as a result of adaptation. Finally, I show in additional experiments that listeners adapt to multiple speakers and that adaptation behavior is modulated by non-linguistic contextual factors such as the speaker's mood. This work has implications for both semantic theories of uncertainty expressions and psycholinguistic theories of adaptation: it highlights the need for dynamic semantic representations and suggests that listeners integrate their general linguistic knowledge with speaker-specific experiences to arrive at more precise interpretations

Language users draw all kinds of inferences concerning the opinions, moods, backgrounds, and social relations of speakers on the basis of what they say, and much of what is conveyed depends only indirectly, if at all, on the literal content of what is said. Though meaning beyond the literal comprises a hefty and potent share of linguistic meaning--as recognized in the traditions of both Gricean pragmatics and meaning-based sociolinguistics--much remains to be uncovered and explained as regards this domain. In this work, I develop a socio-pragmatic framework for understanding meaning beyond the literal, with an eye toward social meaning in particular, and with definite referential phrases as my empirical focus. I apply the framework in two case studies of social meaning, examining the definite article 'the' and demonstratives. In the first case study, I show that, in referring to a group of people, using 'the' ('the Americans') as opposed to a bare plural ('Americans') tends to depict the group as a monolith of which the speaker is not a part. Second, I sharpen the insights of previous research on the social meanings of demonstratives (e.g., Lakoff 1974), and explain why they serve as a useful resource for expressing exclamativity and evaluativity and for promoting a sense of shared perspective and experience between interlocutors. In both cases, I explain the relevant social meanings via the socio-pragmatic framework developed herein: we can understand the social meanings associated with these expressions by examining the expressions' content in the light of contextual (especially social) factors and the content of functionally related alternative expressions. In addition to providing new insights into the social character of English determiners, this work makes the case that social meaning is an indispensable facet of interpretation and use, demonstrates the advantages of pursuing semantic, pragmatic, and sociolinguistic research in tandem, and pushes toward a unifying theory of meaning beyond the literal.

The ability of a machine to communicate with humans has long been associated with the general success of AI. This dates back to Alan Turing's epoch-making work in the early 1950s, which proposes that a machine's intelligence can be tested by how well it, the machine, can fool a human into believing that the machine is a human through dialogue conversations. Despite progress in the field of dialogue learning over the past decades, conventional dialog systems still face a variety of major challenges such as robustness, scalability and domain adaptation: many systems learn generation rules from a minimal set of authored rules or labels on top of handcoded rules or templates, and thus are both expensive and difficult to extend to open-domain scenarios. Meanwhile, dialogue systems have become increasingly complicated: they usually involve building many different complex components separately, rendering them unable to accommodate the large amount of data that we have to date. Recently, the emergence of neural network models the potential to solve many of the problems in dialogue learning that earlier systems cannot tackle: the end-to-end neural frameworks offer the promise of scalability and language-independence, together with the ability to track the dialogue state and then mapping between states and dialogue actions in a way not possible with conventional systems. On the other hand, neural systems bring about new challenges: they tend to output dull and generic responses such as ``I don't know what you are talking about"; they lack a consistent or a coherent persona; they are usually optimized through single-turn conversations and are incapable of handling the long-term success of a conversation; and they are not able to take the advantage of the interactions with humans. This dissertation attempts to tackle these challenges: Contributions are twofold: (1) we address new challenges presented by neural network models in open-domain dialogue generation systems, which includes (a) using mutual information to avoid dull and generic responses; (b) addressing user consistency issues to avoid inconsistent responses generated by the same user; (c) developing reinforcement learning methods to foster the long-term success of conversations; and (d) using adversarial learning methods to push machines to generate responses that are indistinguishable from human-generated responses; (2) we develop interactive question-answering dialogue systems by (a) giving the agent the ability to ask questions and (b) training a conversation agent through interactions with humans in an online fashion, where a bot improves through communicating with humans and learning from the mistakes that it makes.

In this thesis, I study the variables that affect the selection between semantically similar adjectives in English, such as ``big''/``large'', ``high''/``tall'', and ``fast''/``quick''. Currently, there is no general model for predicting the selection between adjectives that compete on a given meaning. The proposals that do exist focus on specific pairs of adjectives rather than the selection between them across the board. I attempt to bridge this gap by integrating models of lexical competition with theories of adjective semantics. I propose a novel distinction between an indicative property of an adjective and its focal dimensions. The former is the core meaning of the adjective, and it is an abstract property with undetermined variables. The latter are possible value assignments for these undetermined variables, which can saturate the indicative property in different ways depending on which ones are selected. I show how these two meaning components interact to yield the selectional preferences expressed by adjectives. Motivated by this theoretical analysis, as well as well-known approaches to pragmatics, I raise the hypothesis that semantically similar adjectives should express a distributional pattern in which the rarer one conveys a more specialized meaning than the more frequent one. I operationalize this hypothesis using the Rational Speech Acts framework, and demonstrate through a series of quantitative studies how it predicts the empirical distribution of adjectives

Human language text plays a pivotal role in medicine. We use text to represent and store our biomedical knowledge, to communicate clinical findings, and to document various forms of medical data as well as healthcare outcomes. While deep language understanding techniques based on neural representation learning have fundamentally advanced our ability to process human language, can we leverage this advancement to transform our ability to understand, generate and utilize medical text? If so, how can we achieve this goal? This dissertation aims to provide answers to these questions from three distinct perspectives. We first focus on a common form of medical text, biomedical scientific text, and study the long-standing challenge of extracting structured relational knowledge from this text. To handle the long textual context where biomedical relations are commonly found, we introduce a novel linguistically-motivated neural architecture that learns to represent a relation by exploiting the syntactic structure of a sentence. We show that this model not only demonstrates robust performance for biomedical relation extraction, but also achieves a new state of the art on relation extraction over general-domain text. In the second part of this work, we focus on a different form of medical text, clinical report text, and more specifically, the radiology report text commonly used to describe medical imaging studies. We study the challenging problem of compressing long, detailed radiology reports into more succinct summary text. We demonstrate how a neural sequence-to-sequence model that is tailored to the structure of radiology reports can learn to generate fluent summaries with substantial clinical validity. We further present a reinforcement learning-based method that optimizes this system for correctness, a crucial metric in medicine. Our system has the potential of saving doctors from repetitive labor and improving clinical communications. Finally, we connect text and image modalities in medicine, by addressing the challenge of transferring the knowledge that we learn from text understanding to understanding medical images. We present a novel method for improving medical image understanding by jointly modeling text and images in an unsupervised, contrastive manner. By leveraging the knowledge encoded in text, our method reduces the amount of labeled data needed for medical image understanding by an order of magnitude. Altogether, our studies demonstrate the great potential that deep language understanding and generation has in transforming medicine

Many profound insights from biomedical research and clinical practice remain hidden within the unstructured text of scientific articles and electronic medical records. Extracting structured information from biomedical text could dramatically accelerate the pace of biomedical research, but due to the high variability of natural language, it hinges on our ability to recognize when different-looking statements are saying the same thing. Unfortunately, attempts to address this problem in the biomedical domain usually involve structured lexicons and ontologies, which are expensive and time- consuming to produce. In recent years, a subdomain of natural language processing called distributional semantics has approached normalization in a different way: by learning mathematical representations of words, phrases, and relationships based on their usage patterns in large corpora. These methods can detect that two different strings are semantically related based on how they are used in context, and require little or no human effort. This dissertation illustrates how distributional approaches can be applied to several important biomedical text mining tasks, including gene, drug and disease name normalization, ontology building, and the construction of a structured radiology lexicon from clinical notes. I describe a novel distributional algorithm (EBC) for extracting relationships among biomedical entities, such as chemicals, genes and diseases, and show how it can be applied to learn the structure of chemical-gene, chemical-disease, gene-disease, and gene-gene relationships from contextual usage patterns. Finally, I apply distributional relationship extraction to two inferential tasks: curating pharmacogenomic pathways, and uncovering the mechanisms behind drug-drug interactions.

Discourse particles provide important signals in conversation, by helping speakers and hearers coordinate on the course of an interaction. Therefore, a precise understanding of discourse particles will provide new insights into the pragmatics of conversation. In this thesis, I will present a framework based on questions under discussion that allows us to capture the key information-theoretic structures in conversation that seem to affect the use of discourse particles: the presence or absence of presuppositions, the issues guiding a conversation, and how interlocutors move between these issues. I present two case studies of German discourse particles that highlight central aspects of the QUD framework: überhaupt and doch. These raise a challenge found in particle systems in many languages: lexicalized focus. Many languages possess particles that can occur with or without focus, and the meanings associated with the unfocused and focused variants are often very different. Since intonation can have discourse-managing functions similar to that of discourse particles, the effect of having or lacking focus marking directly on a particle is different from the effect of focus on regular content words. I will identify patterns that allow us to systematically distinguish the meanings of focused and unfocused particles in a focused/unfocused pair. This serves as a stepping stone towards understanding the interplay of grammar, intonation, and interaction.

This dissertation explores the interplay of conventional and interactional factors in the interpretation of natural language utterances. It develops a formal framework, dynamic pragmatics, in which pragmatic inferences arise as contextual entailments in a dynamic system in which information states are updated with information about the occurrence of utterance events (in contrast to dynamic semantics, where information states are updated with the content of linguistic expressions). In this way, the framework is able to faithfully model Gricean pragmatic inference as interlocutors' reasoning about each other's utterance choices. Linguistic utterances are analyzed as having essential facts of two distinct types: Epistemic effects (i.e., effects on the information states of the interlocutors) and normative effects (i.e., effects on the interlocutors' commitments). The latter effects are carried by extra-compositional, normative conventions of use that mediate the form-force mapping; the former arise largely due to the interlocutors' presumptions about each other's beliefs, preferences, and method of determining which (utterance) actions are best (i.e., practical reasoning). The framework of dynamic pragmatics allows us to consistently take a thoroughly Gricean perspective on language use, and allows us to explore how the interpretation of an utterance arises through the interplay of sentential force, content, and context. At the same time, the framework of dynamic pragmatics sheds a new light on the nature of conversational implicature, and language use in general.

This dissertation investigates the semantic foundations of nominal countability. Standard accounts are typically concerned with a binary distinction between countable words (`dog'/`dogs') and non-countable words (`water'). This dissertation examines this issue from the perspective of languages with richer grammatical number systems. I develop a typological generalization that countability is a scalar phenomenon and propose new techniques to formally model these facets of nominal semantics by augmenting standard mereological accounts with topological relations. Languages such as Welsh or Maltese grammatically recognize what I call aggregate nouns--nouns which designate entities that habitually come together, such as insects (ants) or granular substances (sand). These nouns are grammatically distinct from both non-countable nouns and countable nouns with a singular/plural contrast, instead they display a collective/singulative contrast. These grammatical number systems vividly demonstrate how a binary countable/non-countable distinction oversimplifies the typological space. I argue from the data from Welsh and Maltese, and even more complex fieldwork data from the Gur language Dagaare, that countability is a scalar phenomenon. I propose that the morphosyntactic organization of grammatical number systems reflects the semantic organization of noun types according to the degree of individuation of their referents. Nouns of different types are individuated to different degrees and can accordingly be ordered along a scale of individuation: substances < granular aggregates < collectives < individual entities. Noun types which are less individuated are on the lower end of the scale and are cross-linguistically less likely to signal grammatical number, while the converse holds for highly individuated noun types. Understanding morphosyntactic number categories in light of a scale of individuation avoids the difficulties binary accounts face, since languages may divide up the scale of individuation into any number of classes and at different points. For instance, languages with a collective/singulative recognize a grammatical number category corresponding to the middle region of the scale. At the same time, the proposal provides a predictive framework for how grammatical number systems are organized: the contrasts being made are common across languages, and, as a corollary, the endpoints of the scale (substances and individual entities) are predicted to be stable across languages. I show that this view of countability also answers many of the standard criticisms of accounts where a noun's meaning determines its grammatical behavior with respect to number marking. I explore the implications of this broader typological view for formal semantic treatments of countability. Standard mereological accounts turn out to be not sufficiently expressive to model the aggregate nouns nor the grammatical number systems which distinguish them. I enrich the standard mereology framework with topological connection relations, resulting in the more expressive ``mereotopology". Through using different connection relations, this framework is able to represent aggregate nouns and the ways in which entities may come together. Consequently, this framework is able to deliver analyses of particular grammatical number systems, such as Welsh. In addition, this more expressive framework resolves several recalcitrant problems noted for many treatments of countability, such as the ``minimal parts" problem discussed in relation to nouns such as `sand' or `furniture' which, while non-countable, still have minimal pieces.

Extant research often models ambiguity — that one object can have multiple interpretations — as "noise" or "optional" (e.g. McMahan and Evans 2018, Reagans and Zuckerman 2008). This stance is in tension with evidence that ambiguity is common, exists to varying degrees, and is consequential. Ambiguity is therefore better thought of as "more or less" rather than "yes or no". Predictably, literatures that seek to bracket ambiguity find themselves without the analytical tools to specify core theoretical claims and devise measurement strategies. This dissertation works towards taking ambiguity seriously by introducing microfoundations for its analytical specification and a measurement strategy at the level of the word. To demonstrate the benefits and broader ramifications of this orientation, these theoretical and methodological tools are applied to the "brokerage" literature (Burt 1982) to surface, clarify, and resolve a long-standing contradiction: do brokers merely transmit or actively translate?

Grice (1975) puts forward a view of linguistic meaning in which conversational agents enrich the semantic interpretation of linguistic expressions by recourse to pragmatic reasoning about their interlocutors and world knowledge. As a simple example, on hearing my friend tell me that she read some of War and Peace, I reason that, had she read all of it, she would have said as much, and accordingly that she read only part. It turns out that this perspective is well suited to a probabilistic formalization. In these terms, linguistic meaning is fully characterized by a joint probability distribution P(W; U) between states of the world W and linguistic expressions U. The Gricean perspective described above corresponds to a factoring of this enormously complex distribution into a semantics [[u]](w) : U -> (W -> {0, 1}, world knowledge P(W) and a pair of agents which reason about each other on the assumption that both are cooperative and have access to a commonly known semantics. This third component, of back and forth reasoning between agents, originates in work in game-theory (Franke, 2009; Lewis, 1969) and has been formalized in probabilistic terms by a class of models often collectively referred to as the Rational Speech Acts (RSA) framework (Frank and Goodman, 2012). By allowing for the construction of models which explain in precise terms how Gricean pressures like informativity and relevance interact with a semantics, this framework allows us to take an intuitive theory and explore its predictions beyond the limits of intuition. But it should be more than a theoretical tool. To the extent that its characterization of meaning is correct, it should allow for the construction of computational systems capable of reproducing the dynamics of opendomain natural language. For instance, on the assumption that humans produce language pragmatically, one would expect systems which generate natural language to most faithfully reproduce human behavior when aiming to be not only truthful, but also informative to a hypothetical interlocutor. Likewise, systems which interpret language in a human-like way should perform best when they model language as being generated by an informative speaker. Despite this, standard approaches to many natural language processing (NLP) tasks, like image captioning (Farhadi et al., 2010; Vinyals et al., 2015), translation (Brown et al., 1990; Bahdanau et al., 2014) and metaphor interpretation (Shutova et al., 2013), only incorporate pragmatic reasoning implicitly (in the sense that a supervised model trained on human data may learn to replicate pragmatic behavior). The approach of this dissertation is to take models which capture dynamics of pragmatic language use and apply them to open-domain settings. In this respect, my work builds on research in this vein for referential expression generation (Monroe and Potts, 2015; Andreas and Klein, 2016a), image captioning (Vedantam et al., 2017) and instruction following (Fried et al., 2017), as well as work using neural networks as generative models in Bayesian cognitive architectures (Wu et al., 2015; Liu et al., 2018). The content of the dissertation divides into two parts. The first (chapter 2) focuses on the interpretation of language (particularly non-literal language) using a model of non-literal language previously applied to hyperbole and metaphor interpretation in a setting with a hand-specified and idealized semantics. Here, the goal is to instantiate the same model, but with a semantics derived from a vector space model of word meaning. In this setting, the model remains unchanged, but states are points in an abstract word embedding space - a central computational linguistic representation of meaning (Mikolov et al., 2013; Pennington et al., 2014). The core idea here is that points in the space can be viewed as a continuous analogue of possible worlds, and that linear projections of a vector space are a natural way to represent the aspect of the world that is relevant in a conversation. The second part of the dissertation (chapters 3 and 4) focuses on the production of language, in settings where the length of utterances (and consequently the set of all possible utterances) is unbounded. The core idea here is that pragmatic reasoning can take place incrementally, that is, midway through the saying or hearing of an utterance. This incremental approach is applied to neural language generation tasks, producing informative image captions and translations. The result of these investigations is far from a complete picture, but nevertheless a substantial step towards Bayesian models of semantics and pragmatics which can handle the full richness of natural language, and by doing so provide both explanatory models of meaning and computational systems for producing and interpreting language

When a person says something that has multiple possible interpretations, which interpretation stands out as the most likely intended meaning often depends on context outside the utterance itself: salient objects in the environment, utterances the speaker could have chosen but didn't, common-sense knowledge, etc. Systematically predicting these contextual effects is a major unsolved problem in computational natural language understanding. A recently-developed framework, known in cognitive science as the rational speech acts (RSA) model, proposes that speaker and listener reason probabilistically about each other's goals and private knowledge to produce interpretations that differ from literal meanings. The framework has shown promising experimental results in predicting a wide variety of previously hard-to-model contextual effects. This dissertation describes a variety of methods combining RSA approaches to context modeling with machine learning methods of language understanding and production. Learning meanings of utterances from examples avoids the need to build an impractically large, brittle lexicon, and having models of both speaker and listener also provides a way to reduce the search space by sampling likely subsets of possible utterances and meanings. Using recently-collected corpora of human utterances in simple language games, I show that a combination of RSA and machine learning yields more human-like models of utterances and interpretations than straightforward machine learning classifiers. Furthermore, the RSA insight relating the listener and speaker roles enables the use of a generation model to improve understanding, as well as suggesting a new way to evaluate natural language generation systems in terms of an understanding task.

People are extremely skilled at understanding and reasoning about the emotions of those around them---what I term Affective Cognition. I propose that people have a rich intuitive theory of emotions, which comprises a structured collection of emotion concepts and the causal relationships between these emotions and (i) events and mental states that 'cause' emotions, as well as (ii) behaviors that are 'caused' by emotions. Affective cognition can thus be understood as domain-general inference applied to this intuitive theory of emotions. In particular, this framework allows the generation of mathematically-principled, quantitative predictions about affective cognition that can be tested through experiments. In this dissertation, I outline a computational, ideal-observer approach that models affective cognition as optimal Bayesian reasoning. I show that this model predicts how human participants make inferences about unseen outcomes from observed emotions. In particular, the model also offers a solution to an age-old debate: how do people infer someone's emotions based on multiple, potentially conflicting cues? The proposed solution to emotional cue integration under this Bayesian framework is joint inference given multiple cues; this inference automatically weighs each cue according to their reliability in predicting emotions. I show that this model accurately tracks human participants' cue integration across a series of experiments, suggesting that this approach provides a promising description of human affective cognition. Affective cognition is inherently social, and people's emotion judgments are affected by their relationship with who they are making judgments about. Across several experiments, I investigated the effect of perceived psychological distance---how similar people were to others that they were making emotion judgments about---on affective cognition. I found that psychological distance biases emotion judgments in two ways: Increasing psychological distance causes people to judge others to feel more negative emotions and less positive emotions; and it causes people to weigh the emotion-relevant features of the situation context more in their emotion judgments. I term this bias in affective cognition Contextualized Self-Enhancement. This work lays the foundation for future work to incorporate psychological distance and other social factors into computational models of affective cognition. I will also describe work that aims to computationally model affective cognition in naturalistic contexts: specifically, modeling how people reason about the emotions of others spontaneously describing past life events. I describe the collection of a large corpus of videos of participants ("targets") describing personally-relevant emotional events in their lives. These unscripted self-disclosures provide a rich, multimodal source of emotional information: facial expressions, acoustic cues (pitch/prosody), as well as the linguistic information in the content of their narratives. I outline a computational model that can predict observers' ratings of targets' emotional valence over the course of these videos. This work will have important implications for understanding affective cognition in real-life contexts, as well as to building computers and robots that can "understand" their users' emotions. In summary, this dissertation examines human reasoning about emotion using computational modeling and behavioral experiments. This theoretically-grounded approach provides a productive research framework, with links to many other areas of psychology, and will enrich our understanding of intuitive human reasoning and of human emotions more broadly. Furthermore, this approach holds great promise for many applications, such as to modeling psychopathology and advancing affective computing technology.

The last few years have seen many striking successes from artificial neural network models on hard natural language processing tasks. These models replace complex hand-engineered systems for extracting and representing the meanings of sentences with learned functions that construct and use their own internal vector-based representations. Though these learned representations are effective in many domains, they aren't interpretable in familiar terms and their ability to capture the full range of meanings expressible in language is not yet well understood. In this dissertation, I argue that neural network models are capable of learning to represent and reason with the meanings of sentences to a substantial extent. First, I use entailment experiments over artificial languages to show that existing models can learn to reason logically over clean language-like data. I then present a large new corpus of entailments in English and use experiments on that corpus to show that these abilities extend to natural language as well. Finally, I introduce a new model that uses the semantic principle of compositionality to more efficiently and more effectively learn language from large volumes of data.

In modern healthcare, there is no accepted system for diagnosing disorders of mental function based on the altered brain systems that cause them. In neuroscience, there is no consensus on what a brain system is. This thesis takes a data-driven approach to mapping domains of brain function across the human neuroimaging literature. Neural circuits and associated mental functions were mapped from the brain coordinate data and full texts of nearly 20,000 studies. The resulting domains characterize several novel brain circuits that are absent from the conceptually dominant expert-determined frameworks in the field. Having engineered a framework for brain function, the next aim was to interrogate whether its domains adhere to basic organizational principles. Are the domains reproducible, meaning that their circuits and functions predict one another in held-out articles? Are they modular, partitioning the literature into homogeneous and separable subsets? Are they generalizable, serving as prototypes of circuits and functions observed in single studies? It is demonstrated here how competing knowledge frameworks can be evaluated against a common set of standards. The data-driven framework is consistently seen to offer leading performance. The second aim of the thesis was to apply the data-driven framework to predict clinical outcomes. The notes from over 20 million healthcare visits were rated by semantic similarity to domains, and these dimensional phenotypes were found to outperform binary diagnoses in predicting future psychotropic prescriptions, hospital visits, and all-cause mortality. Though the data-driven framework was not designed to have clinical relevance, its prognostic performance is on par with frameworks crafted specifically for psychiatric applications. The third and final aim of the thesis was to systematically quantify discrepancies between article texts and brain coordinate data. Computational approaches to ontology engineering are expected to mitigate certain forms of bias, while others may be inherent to the neuroimaging literature. The findings point to the influence of the publication incentive structure on how neuroimaging data are interpreted by neuroscientists. The collective results illustrate the scientific, clinical, meta-scientific interest in engineering a data-driven framework for brain function. They suggest a potential for biological knowledge frameworks to go beyond guiding diagnosis and to generate the prognoses that have long eluded the standard of care for disorders of brain function

Intonation, in particular, terminal contours, interacts with morphosyntactic features of clause-types (declaratives, interrogatives, imperatives, etc.) to help determine the speech act of the utterance and generate complex additional inferences about the context and the speaker. This dissertation addresses the question of how this comes about, focusing on a particular tune + clause-type combination, namely, rising declaratives of American English. English rising declaratives have been associated with a wide range of seemingly disparate meanings. They may be used as tentative assertions (the so-called 'uptalk' uses, often accompanied by social stigma), and they may be used as biased questions. In the latter case, they can sometimes convey positive epistemic bias of the speaker, but other times may convey negative epistemic bias instead. They often convey particular interactional and social meanings, like speaker politeness, but may also convey opposite social meanings, like speaker annoyance or exasperation. Characterizing the core, conventional effect of rising declaratives that crosscuts all of these varied uses has been a challenge. This dissertation presents a series of experimental studies and a theoretical analysis that reconcile these disparate, potentially conflicting observations that have been made about English rising declaratives data. It establishes the existence of two distinct types of rising declaratives, and derives additional, enriched meanings from the conventional effects of these two clause-types and their interactions with context and pragmatic reasoning.

The topic of this dissertation is the famous principle of compositionality, stating that the meanings of complex expressions are determined by the meanings of their parts and how the parts are put together. The title belies its intent: rather than advanc- ing a central thesis, it instead asks and provides answers to new questions about compositionality. In particular, these questions arise by shifting from viewing com- positionality as a property of symbolic systems -- where what I call 'status questions' are naturally discussed -- to viewing it from a procedural perspective, as operative in processes of production and interpretation. The questions the dissertation asks arise at successively narrower levels of scale. At the level of our species, I ask: why are natural languages compositional in the first place? At the level of small conversations: what role does compositionality play in the broader theory of communication? And at the level of an individual language speaker: what is the algorithmic interpretation of compositionality and what demands on complexity does it impose? In the pursuit of answers to these questions, a wide variety of methods are employed, from simulations of signaling games, to logic and formal semantics, to theoretical computer science, as appropriately called for.