Most work on semantic adaptation has focused on comprehenders’ ability to adapt to the production probabilities associated with specific speakers. Here we ask whether semantic features of predicates modulate adaptation. We investigate the relationship between vagueness and adaptation through the English vague quantifiers few and many. Results from two semantic adaptation studies show that 1) many is vaguer than few; 2) the vaguer quantifier many does not adapt, whereas few adapts both when directly primed with few, and when primed indirectly with its antonym many. We argue that there is an inverse relation between adaptation and the degree of vagueness displayed by the quantifier at baseline. The resistance of vaguer predicates to adapt suggests that highly vague meanings are more resilient to change than more precise ones.
People use language to express their thoughts. It is far from clear, however, whether the format of these thoughts (conceptual and perceptual representations) is the same as the semantic representations of natural language. I address this question focusing on two puzzles: 1) how thematic roles as encoded in morphosyntax relate to event role concepts, and 2) how verb argument structure relates to the wide swath of meaning associated with verbs (often called ‘encyclopedic knowledge’). Drawing on experimental studies in English, Basque, Hindi, Spanish, and Chinese, I argue that semantic structures align with conceptual representations in some instances but diverge in others.
Developing accurate and informative representations of events described in documents is a longstanding challenge in NLP. Templates are a classic answer to this challenge, and are especially popular in the field of information extraction. A template represents a complex event of a pre-defined type and is associated with a set of slots that characterize the roles participants may play in that event. This talk will present some recent models for the task of template filling — generating appropriate templates given a document — as well as some of the persistent difficulties posed by the task itself and its evaluation.
This talk addresses the current limitations of existing document-level event-linking approaches, which are often limited by the domains covered by their ontologies. In order to mitigate this, we are developing a broad-coverage cross-document event linking and role extraction corpus, supported by the Framenet ontology. This resource will enable us to link events across multiple documents and extract the roles played by various entities in these events, thus providing a more comprehensive understanding of complex events. We will discuss the methodology used in building this corpus. The findings from this research are expected to significantly enhance our ability to analyze and interpret events across a wide range of domains and to overcome the limitations of existing approaches in document-level event linking.
The burgeoning growth of large language models has brought renewed attention to the need for hybrid dialogue systems that are capable of delibrate conversation driven by goals and knowledge, yet also allowing for robust expectation-driven behavior. In this talk, I will present a dialogue management framework that uses dialogue schemas – structured representations of prototypical dialogue events, expressed in a type-coherent logical form – to flexibly guide conversation. I will discuss a couple applications of this dialogue framework in creating virtual avatars across multiple domains. Finally, I will discuss ongoing work on modelling intensional states relevant to dialogue using crowdsourced human judgments, as well as preliminary work on leveraging large language models for schema-based response generation.
Code-switching happens when a speaker alternates between one language and another. We study how best to model the cognitive process of written code-switching, distinguishing between the Matrix Language Frame (Joshi, 1982, Myers-Scotton, 2002) and the Dual Path Model (Tsoukala et al., 2020). The Matrix Language Frame models code-switching as a main “matrix” language that is supplemented by a secondary “embedded” language, while the Dual Path Model models code-switching as a bi-directional process where speakers access both languages in parallel, producing the most salient continuation at each step. By comparing patterns of code-switching in online posts to monolingual texts, we find support for the Dual Path Model.
In this talk, we will explore the behavior process of creating summaries for expository texts by native English speakers: (1) We examined how people apply different levels of linguistic compression to texts to highlight important events and entities with analyses of syntactic and semantic structures of the summaries; (2) We investigated whether such linguistic compressions are useful and effective for readers to understand the core events through a human study. We will also discuss whether the compression patterns can be learned by computational models with large language models.