Neural network approaches to the study of word learning
Abstract/Contents
- Abstract
- I present three approaches to using neural network models for studying word learning in children. In the process, I show what we can conclude from these different approaches about how words are learnt by humans as well as by neural networks. The first approach takes neural networks to be cognitive models, allowing for stronger conclusion about the connection between model and human word learning outcomes. I show that the predictability of words in the linguistic contexts children are exposed to is a good predictor of the age at which children will begin to produce these words themselves. Additionally, I demonstrate that language models share similarities with children when it comes to the order in which they learn new words, but only if these words are predicates or function words. The second approach considers neural networks to be independent learners from humans and uses them as a "proof of concept" tool. These model learners can exemplify what behaviors are in practice learnable rather than inherent. Here, I show that the meanings of logical operators like "and" and "or" are in fact learnable by models that are exposed to visually grounded language, and therefore, could also be learnt without prior biases by human learners, supporting existing usage-based theories of their acquisition. I also find that these visually-grounded models can learn the meanings of other function words, favoring gradient semantic representations over threshold based semantics for comparative quantifiers and spatial prepositions. The third and final approach, like the second, considers models as independent learners, but instead of using them to differentiate between existing proposals, it considers them as a tool to generate novel hypotheses about what drives language learning in both models and children. We can establish novel causal relationships about what allows models, and by extension possibly humans, to learn a given outcome. I show that neural agents exposed to social interaction in addition to visual and linguistic input can acquire a shape bias when learning new lexical categories, and furthermore, that this bias follows from efficient communication strategies. I propose that the shape bias seen in humans could also follow from these same communication strategies.
Description
| Type of resource | text |
|---|---|
| Form | electronic resource; remote; computer; online resource |
| Extent | 1 online resource. |
| Place | California |
| Place | [Stanford, California] |
| Publisher | [Stanford University] |
| Copyright date | 2022; ©2022 |
| Publication date | 2022; 2022 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Portelance, Eva |
|---|---|
| Degree supervisor | Frank, Michael C, (Professor of human biology) |
| Degree supervisor | Jurafsky, Dan, 1962- |
| Thesis advisor | Frank, Michael C, (Professor of human biology) |
| Thesis advisor | Jurafsky, Dan, 1962- |
| Thesis advisor | Degen, Judith |
| Thesis advisor | Laroche, Romain |
| Degree committee member | Degen, Judith |
| Degree committee member | Laroche, Romain |
| Associated with | Stanford University, Department of Linguistics |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Eva Portelance. |
|---|---|
| Note | Submitted to the Department of Linguistics. |
| Thesis | Thesis Ph.D. Stanford University 2022. |
| Location | https://purl.stanford.edu/vc984hq1552 |
Access conditions
- Copyright
- © 2022 by Eva Portelance
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...