Semantics Relations Research

$1750 (summer '19) & $3500 (summer '20) fellowship

From the start of my freshman Spring (02/2019) to the end of my junior Fall (12/2020), the main focus of my research projects have been on semantic relations. I’ve studied the taxonomic and thematic (or associative relations in some literature) word relations from both developmental and computational aspects.

Understanding how children acquire structured semantic networks is critical both for theoretical accounts of knowledge acquisition and for applied endeavors such as reducing inequalities in the knowledge of school-relevant domains, and studying the differential effects of different semantic relations on word production help fine-tune the models and hypotheses on cognitive architecture of language.

During my freshman and sophomore years, I primarily worked with Dr. Catarina Vales and Dr. Anna Fisher in the Cognitive Development Lab to study the correlation between different patterns of linguistic inputs and children’s semantic structure. The behavioral data was collected from children of 4-6 years old using the Spatial Arrangement Method (SpAM). The statistical measures we’ve used to capture the thematic and taxonomic relations include PMI, LSA, and GloVe, and the corpora we’ve tried to calculate or analyze these measures include Childes, Wiki, Common Crawl, and TASA.

Gephi was used to visualize the word net of different word relations1. R library ggplot was also used to aid the stimulus selection process.

In 2019 summer, I was awarded a stipend to conduct research through the summer. Me and with another student were compensated from the $3500 Ireland undergraduate research fund.

In 2020 summer, I received a $3500 Summer Undergraduate Research Fellowship (SURF) to work on the proposal “The contribution of taxonomic and associative language statistics to the development of structured semantic networks” and presented at lab meetings, Undergraduate Research Meet-Up and Meeting of the Minds Undergraduate Research Symposium (MoM).

The poster @ 2021 MoM
The video presentation @ 2021 MoM

During my junior Fall, I worked in Dr. Bonnie Nozari’s lab on a similar project to study the differential effects of taxonomic and associative/thematic word relations in picture naming task and the typing modality. Meanwhile, in my Parallel Distributed Processing class (85-419/719) taught by Dr. David Plaut, I implemented (using Lens), presented, and wrote a short paper on a simulation model on this task using a fully recurrent network.

I’ve practiced and gained a lot of skills during these research experiences, including Python (NLTK library), R, Gephi, jspsych javaScript, HTML, CSS, Lens, and I also had some exposures to those tools that host online psychology experiment like Qualtrics, psiTurk and Amazon Mechanical Turk (MTurk).

Some selected stimulus of the typing picture-naming experiment.

This is the fully recurrent network model architecture I used for PDP’s final paper.

  1. Regarding the abbreviations, for example, HGLL denotes High taxonomic (Global) and Low associative/thematic (Local) similarity.