Week of Feb 3

Harvard at TISLR

The 15th Theoretical Issues In Sign Language Research (TISLR) was held January 14-17 in Addis Ababa, Ethiopia, with several presentations connected to Harvard Linguistics. 

  • PhD student Natasha Thalluri presented a poster on her work with Kathryn Davidson on "Universality and variation in sign language comparatives",
  • Masashi Tamura (Gallaudet PhD student, M&M Lab RA) presented a poster on his work with postdoc Nozomi Tomita and Kathryn Davidson on "Negative possessive and existential sentences in Japanese Sign Language (JSL)",
  • Nozomi Tomita presented a poster on "Beyond the Paradox: Understanding ASL’s Endangerment through Assessment Frameworks", and
  • Marianthi Koraka (Goettingen, former visiting graduate student) presented a poster on "Negative imperative speech acts in two sign languages," with Markus Steinbach (Goettingen) and Kathryn Davidson.

In especially exciting news, Masashi Tamura won a conference-wide award for "best student presentation" for his presentation on Negative possessive and existential sentences- congratulations to all! 

 

LangCog

The next LangCog meeting of the semester will be on Tuesday, February 4th from 5:30-7:00pm! It will take place in William James Hall, Room #1550. Our speaker next week is Josh Hartshorne, and the title and abstract of his talk can be found below. Food will be available at the meeting, and you can find the schedule for the remainder of the semester on our website.

Title: Grounding language in thought with Probable World Semantics

Abstract: Large Language Models (LLMs) are trained on language-learning as a problem of pattern identification. A long-standing theoretical approach to human language learning treats it as a problem of finding a decryption algorithm. Intuitively, language is a mechanism for getting thoughts from one head into another --- thoughts that exist independently of language itself. While this approach in principle addresses many theoretical problems, it has been frustratingly difficult to implement in computational models. Over the last decade, however, advances in formulating generative models of cognition have put such models ... if not in reach, then at least within sight.

This talk is divided into roughly three equal parts. The first part lays out the theoretical foundation to Probable World Semantics models (a Bayesian implementation of Possible World Semantics), our alternative to LLMs. The second part presents a series of experiments involving Winograd Schema, comparing Probable World Semantics models to several classic theories as well as LLMs, showing that Probable World Semantics addresses key challenges the others struggle with. (LLM's "defeat of the Winograd Schema challenge" turns out to be overstated.) In the final part, I talk about my attempts to extend this work by using intuitive mechanics as a testbed for experiment and theory. This work is still very much in early stages, and I am very interested in feedback from this community.

Share on: