Skip to content Skip to navigation

FriSem - FYPs: Katherine Herman/Jon Walters, Department of Psychology, Stanford University

April 21, 2017 -
3:15pm to 4:30pm
Jordan Hall (Building 420), Room 050
Katherine's Title: Learning word meanings from reference games: integrating semantics and pragmatics
 
Katherine's Abstract: Often, semantics and pragmatics are modeled as separate processes: models like convolutional neural networks that learn visually grounded word meanings generally do not take pragmatic information into account, while pragmatics models such as the Rational Speech Act model (RSA) typically take word meanings as given. We explored the idea that semantics and pragmatics should in fact interact in rich ways: semantic models that account for pragmatic intent should learn word meanings more quickly, and pragmatic models equipped with continuous, learned semantics should be able to capture a wider range of phenomena.
 
I will present a family of models that combine a probabilistic pragmatics component (RSA) with a neural network vision model. The neural network allows the models to flexibly and efficiently learn literal word meanings, and RSA lets them leverage pragmatic reasoning to produce and comprehend words effectively in context. We evaluate these models on the reference game dataset of Graf et al. (2016), and compare to a set of baseline models that do not include explicit pragmatics and/or consider context.
 
Jon's Title: Forming event representations from temporal community structure
 

Jon's Abstract: I will discuss two studies (Schapiro et al., 2013; Karuza et al., 2017) that investigated how humans form event representations and learn about higher-order temporal structure based on viewing sequences of images produced by different generative processes (i.e., walk types) over a graph bearing a community structure. After exposure to a sequence generated by a random walk over the graph, participants showed evidence of perceiving event boundaries upon entering and exiting communities. As the latter study failed to replicate this finding, we are launching our own replication experiment on MTurk. I will present progress on this replication and ways of building on this work that connect to active learning contexts.FYPs: Katherine Herman/Jon Walters, Department of Psychology, Stanford University

Event Sponsor: 
Department of Psychology
Contact Email: 
rxdh@stanford.edu