Language, Logic, and Machines (LLaMa) Group
Humans use language to transmit information and to support reasoning. If you ask me how many of the tacos are left, and I say "some", you will defeasibly infer that it's not the case that all are left. This inference depends both on your beliefs about my intentions, but also facts about the meaning of words like "some" and "all", as well as their relations to each other in the language. Even this simple example shows that we need a theory of natural language semantics that explains how people compute the meanings of expressions based on their parts, but also how those meanings become enriched in context in virtue of how interlocutors reason about each others beliefs, goals, and intentions.
Research in the LLaMa Group aims to address this question by combining tools and insights from formal, logic-based approaches to meaning with tools and insights from computational linguistics and work in computational cognitive modeling. In the ideal, the result are theories that combine the richly structured repesentations we believe underlie natural language meaning with a plausible model of how humans learn and perform computations over these representations that accord with behavioral data.
People
- Faculty
Robert Henderson
- Graduate Students
Patricia Lee
Seongjin Park
Stanley Donahoo
- Undergraduate Students
Colton Flowers
- Alums
Shaun Marie Stienestra, Andrea Maynard, Jianrong Yu
Current Reseach
- The semantics and pragmatics of dogwhistle language.
Joint work with Elin McCready that focuses on dogwhistle language---language that sends one message to an outgroup while at the same time sending a second (often taboo, controversial, or inflammatory) message to an ingroup. We explore in a Bayesian game-theoretic setting, the formal properties of dog whistle language and interactional pressures that govern its use. Robert Henderson will be in Japan in 2019 working on two books on these topics in residence at the Singularity Institute at Aoyama Gakuin University.
- Textual entailment.
Patricia Lee is exploring methods of recognizing semantic entailments of sentences in text, inspired by the Recognising Textual Entailment (RTE) task (Dagan et al., 2005), in the SemEval-2010 conference. More specifically, she is exploring methods that are based on using logical-form semantic representations of sentences derived from the Minimal Recusion Semantics framework (Copestake et al, 2005) to compute entailment relationships between sentences.
- Mathematical properties of Lambek Calculi.
Work with Colton Flowers exploring the metamathematical properties (e.g., completeness) of Lambek Calculi that support rich notion of movement and scope-taking.
- Composition and construal in recursive neural networks.
Joint work with Aaron Steven White and Patrick D. Elliot that is exploring computational models of construal to tease apart what factors constrain the possible relations allowed by a predicate.
- Detecting statements of opinion
Seongjin Park is working on detecting statements of opinion vs. statements of fact in corpora, and comparing performance on this task by different types of neural network models.
- Processing secondary content
Stanley Donahoo is working on the processing of secondary content in slurs and expletives using EEG and other behavioral measures.
Past Reseach
- Partisianship and the political language of bills.
This research was conducted by an undergraduate, Shaun Marie Stienestra, using tools from nlp to model partisan reaction to bills based on purely stylistic aspects of their construction.
Courses & Workshops
- Seminar in "Computational Semantics and Pragmatics".
This course has two prongs: (i) we consider recent work borrowing formal tools from functional programming and programming language design to model phenomena in natural language semantics, and (ii) we explore recent work in the computational modeling of pragmatic behavior, mostly through Bayesian Rational Speaker Theory, but also other approaches.