Student news: Naomi Tachikawa Shapiro

Submitted by Joyce Parvi on

Congratulations to PhD candidate Naomi Tachikawa Shapiro, who was recently awarded the Graduate School Presidential Dissertation Fellowship in the Arts, Humanities, Social Sciences, and Social Professions. The award will provide support as she works towards completion of her dissertation, Crosslingual processing in humans and machines: Case studies in morphosyntax. As she describes her research:

 

“My dissertation lays the groundwork for my long-term research agenda on crosslingual processing: how humans and machines process crosslinguistic features. This line of inquiry draws a distinction between the modifier “crosslinguistic”, meaning across languages, and “crosslingual processing”, which I use to refer specifically to the processing of crosslinguistic features. In my dissertation, I propose methods for juxtaposing crosslingual processing in humans and neural language models (“machines”), with the goals of (i) achieving a more holistic understanding of human language processing and (ii) improving crosslingual processing in machines. Focusing on the level of morphosyntax, I am currently adapting the artificial language learning paradigm to conduct a large-scale analysis of crosslinguistic variation in human crosslingual processing, challenging the presupposition that, because crosslinguistic features serve analogous functionsacross languages, they must be learned and processed similarly by speakers crosslinguistically. 

 

My dissertation pulls from my prior work (Shapiro, Paullada, & Steinert-Threlkeld, 2021) on probing the morphosyntactic properties captured by Google's multilingual BERT (Devlin et al., 2019). After examining 166 linguistic features across 13 languages—Afrikaans, Arabic, Chinese, Croatian, Finnish, Hebrew, Korean, Marathi, Spanish, Slovenian, Tagalog, Turkish, and Yorùbá—we found that BERT is capable of recognizing many linguistic features, but only occasionally uncovers when a property is shared by multiple languages. For instance, BERT connected pronouns in the related languages Hebrew and Arabic, but it did not connect the nominative case inflections in Spanish and Turkish. In other words, BERT sometimes processed the same feature differently across languages. 

 

In short, my dissertation asks whether BERT’s differential encoding of the crosslinguistic status of these features is human-like or unhuman-like. How similarly do speakers process shared morphosyntactic features across languages—how do we even operationalize such comparisons?—and what might drive crosslinguistic similarities and differences in morphosyntactic processing?”

 

Shapiro is advised by Shane Steinert-Threlkeld.

News Topic
Share