Lead a Spanish-English code-switching study from design to analysis. I built a pipeline that computes surprisal for XLM-RoBERTa and Llama 3 on minimal pairs and compares model preferences with human judgments, and I mentor undergraduate research assistants on sentence generation and quality control.
Read moreBuilt Python and R pipelines to clean and analyze repeated-reference bilingual data, compared SBERT sentence embeddings using cosine similarity and t-SNE, and fit mixed-effects models to quantify how conventions form over repeated interactions.
Read moreFine-tune and evaluate a transformer-based ZIPA speech model for keyword search in unseen languages, develop the training and inference pipelines, and validate transcriptions to keep training and evaluation reliable.
Read moreUsing Pew Research Center survey data from 3,796 U.S. adults, we fit nested logistic and linear regression models to test whether income and education predict AI chatbot adoption and frequency of use. Education was the strongest socioeconomic predictor, with college graduates having 2.5 times the odds of adoption. Income mattered only above $90,000, and age was the strongest predictor overall.
Read moreUsing THINGS, a dataset of 4.7 million odd-one-out judgments from 12,340 participants, we compared layer-wise representations from a shallow CNN (AlexNet) and a deep CNN (ResNet-50) with human object representations, using odd-one-out accuracy on 50,000 triplets and RSA against human SPoSE embeddings. We expected the deeper ResNet-50 to align more closely with humans, and alignment to peak in mid-to-late layers. The second prediction held, but the shallower AlexNet outperformed ResNet-50 on both measures.
Read moreLed this study of whether feedback type (screencast vs. written) favors content-related or grammar-related comments. I built and curated the Writing Feedback Corpus (5,397 instructor comments on 169 undergraduate papers), managed annotation and coding in Google Sheets, and ran a random forests analysis in R. Instructor variability outweighed feedback modality. Manuscript submitted to Applied Corpus Linguistics.
Read moreThis study asked whether screencast feedback builds more rapport between instructors and students than written feedback, using independent samples t-tests on the same corpus. I contributed to corpus construction, annotated and coded comments, and wrote the methods and discussion sections. The results challenged assumptions about the benefits of screencast feedback. Manuscript submitted to CALICO Journal.
Read more