Machine Learning Street Talk (MLST) #062 - Dr. Guy Emerson - Linguistics, Distributional Semantics
Feb 3, 2022
Dr. Guy Emerson, a computational linguist at Cambridge, shares insights into distributional semantics and truth-conditional semantics. The conversations delve into the challenges of representing meaning in machine learning, the importance of grounding language in real-world contexts, and the interplay between cognition and linguistics. Emerson critiques traditional linguistic models, emphasizing the need for flexible frameworks. The discussion also touches on Bayesian inference in language, examining how context influences meaning and the complexities of vocabulary like 'heap' and 'tall'.
AI Snips
Chapters
Transcript
Episode notes
Firth's Focus
- Guy Emerson points out that many citing Firth and Harris on distributional semantics haven't read their work.
- Firth focused on analyzing poetry, not large language model training.
Truth-Conditional Distributional Semantics
- Distributional semantics aims to computationally learn word meanings from text, posing challenges in representation and learning.
- Guy Emerson advocates for truth-conditional distributional semantics, grounding meaning in truth values and real-world entities, distinguishing between words and their referents.
Data Size and Human Language Learning
- Guy Emerson argues that massive language models, trained on unrealistic data volumes, don't reflect human language learning.
- He emphasizes the importance of human-like data sizes for modeling human language acquisition.