Research

Please refer to my CV for a more detailed list of my work.

The source of subject bias in reference processing

The concept of subject, like in many other linguistic phenomena, plays an important role in reference production. When choosing a referent to continue a sentence like ‘John chatted with Mary’, speakers are more likely to choose the preceding subject ‘John’ than the object ‘Mary’. At the same time, speakers are also more likely to use pronouns for the subject antecedent than the object antecedent.

My research so far aims to understand the complicated notion of subject by disentangling its role in reference processing. I have probed this question from three perspectives (though all of them are still at a preliminary stage):

  1. What properties contribute to the subject bias in referential form choice?
  • Not topicality:
    Lam, S.Y. and Hwang, H. (2022). How does topicality affect the choice of referential expressions? Evidence from Mandarin. Cognitive Science. https://doi.org/10.1111/cogs.13190
    - This paper shows that the effect of topicality on referential form choice is different from that of subjecthood.
  1. What is the nature of the special status of subject？
  • At least not due to cognitive saliency:
    Lam, S.Y., Momma, S., Dillon, B. (In prep). Subject is not cognitively more salient: Evidence from reference planning.
  1. Is there difference between the subject bias between production and interpretation?
  • Yes, subject bias in pronoun interpretation cannot be fully explained by the subject bias in pronoun use:
    Lam, S. Y., & Hwang, H. (2024). Pronoun interpretation is more subject-biased than expected by the Bayesian Model. Language, Cognition and Neuroscience, 40(2), 270–289. https://doi.org/10.1080/23273798.2024.2427073
  • Also see this HSP2021 poster of a case study about how subject bias of pronoun production fails to account for the subject bias of Mandarin null pronoun interpretation

Replicating psycholinguistics experiments in Large Language Models

I have a side interest in seeing how LLMs behave in linguistic phonemena that have been investigated in human participants. Rather than simply probing LLM’s performance by looking at the next-word probability, I am interested in replicating the full set of procedure implemented in human participants on LLMs. In other words, you can directly compare the evaluations I have done with a corresponding psycholinguistic paper.

Pronoun priming: No, LLMs cannot capture human patterns of pronoun priming at all

  • Lam, S.Y.#, Zeng, Q.#, Zhang, K.#, You. C, Voigt, R. (2023). Large Language Models Are Partially Primed in Pronoun Interpretation. Findings of the 61st Annual Meeting of the Association for Computational Linguistics.(# indicates equal contribution)

Corresponding to: Johnson, E. D., & Arnold, J. E. (2023). The frequency of referential patterns guides pronoun comprehension. Journal of Experimental Psychology: Learning, Memory, and Cognition, 49(8), 1325.

Production-interpretation asymmetry in pronoun processing: Surprisingly, LLMs can do this

  • Lam, S.Y.#, Zeng, Q.#, Wu, J., Voigt, R. (2025). Leveraging Human Production-Interpretation Asymmetries to Test LLM Cognitive Plausibility. Proceedingsofthe63rdAnnualMeetingoftheAssociation for Computational Linguistics. arXiv preprint. (# indicates equal contribution)

Corresponding to: Lam, S. Y., & Hwang, H. (2024). Pronoun interpretation is more subject-biased than expected by the Bayesian Model. Language, Cognition and Neuroscience, 40(2), 270–289. https://doi.org/10.1080/23273798.2024.2427073

Other research

Apart from psycholinguistics, I do have an interest in theoretical linguistics. I mainly investigate phenomena of one of my native languages, i.e., Cantonese.

Syntax

  • Cantonese Possession: My ongoing qualifying paper, supervised by Kyle Johnson and Faruk Akkuş, investigates external possession and possessor dative in Cantonese.
  • Passive ellipsis clause penalty in VP Ellipsis: My HSP2025 poster, co-authored with Satoru Ozaki, investigated the phenomenon of why passive ellipsis (e.g., John read the report, the confession was too) is always less acceptable in VPE. Rather than an information-structural account proposed by previous literature (Poppels & Kehler, 2019), we argue for a syntax-semantic account that reconstructing the displaced patient argument (the confession) worsens the acceptability of the sentence.

Semantics and Pragmatics

  • Discourse particles: Supervised by María Biezma, I am investigating the discourse function of two sentence-final particles, i.e., gwaa3 and laa1 in Cantonese.

Phonology

  • Cantonese Nasal Merger: LSA2024 proceeding on a OT-based analysis on Cantonese nasal merger, i.e., the optional insertion of [ŋ] onset into a no-onset syllable and the optional deletion of [ŋ] from a [ŋ]-onset syllable.