Cara Leong
@craaaa
Followers
261
Following
1K
Media
23
Statuses
177
linguistics phd student @ nyu, machine acquisitionist/trainer of models, monday crossword finisher (she/her)
Joined February 2009
Submissions for the 2025 Workshop on Cognitive Modeling and Computational Linguistics are due Feb. 16 I humbly request your help with spreading the word
📣 We are happy to share that CMCL 2025 will be co-located with NAACL in New Mexico! 👉The call for papers is out https://t.co/KLledcVd5q SAVE THE DATES, and submit your work! ‼️ Paper submission deadline: Feb 16, 2025 🗓 May 3 or 4, 2025: Workshop (TBA) @naacl @naaclmeeting
0
10
23
👶NEW PAPER🪇 Children are better at learning a second language (L2) than adults. In a new paper (led by the awesome Ionut Constantinescu) we ask: 1. "Do LMs also have a 'Critical Period' (CP) for language acquisition?" and 2. "What can LMs tell us about the CP in humans?"
8
35
239
tinlab at Boston University (with a new logo! 🪄) is recruiting PhD students for F25 and/or a postdoc! Our interests include meaning, generalization, evaluation design, and the nature of computation/representation underlying language and cognition, in both humans and machines. ⬇️
2
22
106
🧐🔡🤖 Can LMs/NNs inform CogSci? This question has been (re)visited by many people across decades. @najoungkim and I contribute to this debate by using NN-based LMs to generate novel experimental hypotheses which can then be tested with humans!
2
14
83
Cara is presenting her paper today (poster P1-E-27), asking whether LLMs can simulate expertise effects just by telling the system "you are an expert in birds" or "an expert in dogs". Check it out! https://t.co/HXpyH93xhs
#CogSci2024
Excited to be going to my first @cogsci_soc next week! I'll be presenting a poster (with @LakeBrenden) about whether multimodal LMs behave like human experts, and would love to meet new (and old) friends!
0
3
13
What is cause and effect? What is a “mechanism”? And how do answers to these questions affect interpretability research? 📜 New preprint! 📜 Two key challenges for causal/mechanistic interpretability, and ways forward. To be presented at the mech interp workshop at #ICML2024:
1
27
137
This work is a great example of how un-blackbox-like and linguistically interesting language models can be when the training data is manipulated!
New preprint! How can we test hypotheses about learning that rely on exposure to large amounts of data? No babies no problem: Use language models as models of learning + 🎯targeted modifications 🎯 of language models’ training corpora!
0
3
28
high praise from someone who trained a whopping ninety LMs (read https://t.co/IK9TouMyCA!)
arxiv.org
Language models learn rare syntactic phenomena, but the extent to which this is attributable to generalization vs. memorization is a major open question. To that end, we iteratively trained...
Such a well executed paper! Can totally feel the feeling of training a whopping 125 LMs (but def worth it!)
1
0
5
Excited to be going to my first @cogsci_soc next week! I'll be presenting a poster (with @LakeBrenden) about whether multimodal LMs behave like human experts, and would love to meet new (and old) friends!
3
2
18
@tallinzen is a great mentor; thanks to @jowenpetty @wtimkey8 for listening to early/bad versions of this paper, @kanishkamisra for commiserating over training 125 models, and friends at SCiL 2023 for comments!
0
0
9
When dealing with exposure to linguistic input on the scale of millions of words, targeted corpus modification can be used to systematically explore how small changes to the input affect learning. For more, check out the paper: https://t.co/n8HPak05Kb 12/12
1
0
7
We showed that frequency is a plausible pathway for learning human-like passivization, but targeted corpus modifications can also be used to develop new hypotheses to test on human learners 👦🏽👶🏼 11/
1
0
3
But this semantic intervention had no systematic effect on passivizability. Instead, changes in passivizability varied depending on the verb being modified, potentially because our modifications don't change the distributions of already-polysemous verbs like ‘take’ much. 10/
1
0
6
Another hypothesis is that a verb’s semantics (in particular its ~affectedness) affects its passivizability. We modify the kinds of arguments that a verb takes, e.g. placing the verb ‘last’ next to arguments meant to occur with ‘drop’. 9/
1
0
4
Tweaking a verb’s frequency of occurrence matters to our models: reducing how often the model saw a verb in the passive voice makes it significantly less passivizable (increasing passive drop). 8/
1
0
5
To test this hypothesis, we reduce how often a verb appears in the passive (below, mutating DROP to be more like LAST). If frequency of exposure to the verb in the passive drives how a model learns passivizability, our modification should make the verb less passivizable. 7/
1
0
3
We first test the entrenchment hypothesis, under which learners conclude that a verb cannot appear in a particular context if that verb appears with substantial frequency in other contexts but never in the context in question. 6/
1
0
5
We test two hypotheses for the features of the input our models are using to learn a verb’s passivizability through 🎯targeted 🎯modifications 🎯to our models’ training corpora, following on from some great work by @kanishkamisra @kmahowald @JumeletJ and others! 5/
1
0
7
We collect human & model acceptability judgments on sentences containing (un)passivizable verbs and find that LMs trained on 100M words of English make judgments that are similar to human judgments (r=0.6)! How do these models learn to judge which verbs are passivizable? 4/
1
0
5