Projects

Ongoing Projects

Five active projects extend the XPLAIN program toward stage-aligned, individually adaptive AI scaffolding for real-time conversation. Together they address how proactive AI tools should be calibrated to user-level cognitive dimensions, language background, conversational alignment, and the social perception of AI-assisted speech.

Personalization of Proactive AI Scaffolds

Wizard-of-Oz study (N = 40); manuscript in preparation.

This project moves from what AI scaffolds to provide (XPLAIN’s lexical clarifications, idea/content suggestions, topic summaries) to for whom and when they should be triggered. Building on the baseline XPLAIN study, I designed and ran a Wizard-of-Oz study (N = 40) in which a confederate simulated context-adaptive interventions calibrated to user-level cognitive dimensions — language proficiency, working-memory profile, AI familiarity, and communication style — under different conversational contexts (high vs. low domain familiarity, comprehension- vs. production-heavy turns).

• Goal: Test whether context-adaptive (vs. one-size-fits-all) scaffolds yield measurable gains in efficiency, participation, and inclusivity for L2 speakers, and whether the gains depend on which user dimension is matched to which context.
• Method: Mixed-methods — multimodal behavioral coding of dyadic virtual meetings, post-task surveys, and semi-structured interviews.
• Status: Data collected; full personalization manuscript in preparation.

User Heterogeneity via Latent Class Analysis

Accepted at CogSci 2026 (Rio de Janeiro, Brazil); full paper in preparation.

Using the qualitative interview data from the personalization study as input, this companion analysis asks a theoretical question: does user heterogeneity in AI-MC reduce to a small number of stable types, or does it operate domain by domain?

• Method: Latent class analysis (LCA) on n = 29 fully coded interview transcripts.
• Key finding: LCA identified eight two-class groupings across thematic domains (AI usage and trust, communication strategies, feature-specific evaluations). User profiles in one domain did not predict profiles in another, indicating that adaptation patterns are domain-specific rather than reflecting a global typology.
• Implication: Personalization for conversational AI should adapt across multiple, partially independent dimensions of user experience — extending classical cognitive theories that assume between-individual variation aggregates into stable types.

L1- vs. L2-cued Translation Scaffolds

Wizard-of-Oz study (N = 27); manuscript in preparation.

For non-native (L2) speakers, real-time turn-taking imposes simultaneous demands: comprehend incoming speech, predict appropriate responses, access target-language lexical forms, and articulate under turn-taking constraints. AI mediation adds the further step of interpreting and integrating AI-suggested content. This project asks whether AI scaffolds should be delivered in the user’s L1 or L2 under domain-specific lexical demand.

• Method: Wizard-of-Oz study (N = 27) crossing scaffold-language (L1 vs. L2) with domain familiarity, integrating prediction theory, cognitive load theory, and language-dependent memory theory.
• Theoretical contribution: Operationalizes prediction and cognitive-load theories — developed for face-to-face contexts — for AI-mediated communication, where interpretation and evaluation delay speech planning and production.
• Status: Long work-in-progress paper; CHI poster review version drafted.

Interactive Alignment & Semantic Integration

Manuscript in preparation.

Subjective-satisfaction metrics dominate the evaluation of conversational AI, yet they fail to capture whether AI mediation preserves shared common ground or merely papers over local breakdowns. This project introduces a dual-lens framework for evaluating proactive AI scaffolding in L2 turn-taking:

• Semantic Integration — do users integrate AI suggestions into their own discourse context (paraphrase, build on, contextualize) rather than verbatim copy them? Measured with high-dimensional text-similarity analyses on the XPLAIN Wizard-of-Oz corpus.
• Cognitive Fluency — does offloading prediction yield a measurable reduction in speech disfluencies (filled pauses, repairs, prolonged silences) at theoretically motivated stages of speech production?
• Goal: Identify the “Integration Zone” — the operating regime where algorithmic assistance reduces cognitive load without replacing the human speaker — and connect HCI evaluation back to the psycholinguistic theory of interactive alignment.

Listener Perception of AI-Assisted Speech

Pre-registered on OSF; in data collection (Spring 2026).

The same multimodal traces that allow AI tools to detect when a speaker needs help (disfluency, timing, prosody, visible ease/strain) may also leak to listeners — and listeners may attribute those traces to the speaker rather than to the AI. This project tests the resulting “double bind” for L2 speakers using AI support.

• Design: Pre-registered two-phase listener-perception study. Phase 1 is blind (listeners rate naturalistic clips of L2 speech with and without AI scaffolding without being told AI was involved). Phase 2 is informed (listeners are told some clips were AI-assisted and asked to identify which).
• Outcomes: Implicit detection accuracy of AI-assisted speech; acoustic-cue attribution; competence and authenticity ratings; moderation by listeners’ prior AI knowledge.
• Why it matters: If AI-assisted speech triggers negative competence/authenticity attributions even when listeners cannot consciously detect it, then proactive AI tools that succeed on cognitive metrics may fail on social metrics — a critical design tension for any AI-MC system aimed at non-native speakers.

Past Projects

Sustaining Public Goods via Prosocial Behavior

Citizens & Technology Lab (CAT Lab), Cornell University. PI: J. Nathan Matias. Jan. 2025 – Present (manuscript under review at PNAS).

Lead statistician on parallel pre-registered field experiments across four Wikipedia language communities (Arabic, German, Persian, Polish; N = 15,558 vetted editors), quantifying the causal effect of peer-to-peer gratitude on upstream reciprocity and sustained volunteer participation. Built end-to-end causal-inference pipelines in R (power analyses; intent-to-treat and complier-average causal-effect estimation; multilevel models with cluster-robust SEs; sensitivity analyses; cross-community meta-analytic pooling). Receiving thanks increased contribution time by 11%, two-week retention by 2.2 pp, and outgoing thanks by 61%, with 99.8% sent upstream. See the PNAS manuscript.

Cross-Linguistic Statistical Learning of Language

Cognitive Science of Language Lab, Cornell University. PI: Prof. Morten H. Christiansen. Aug. 2021 – May 2023.

Investigated how high-frequency multi-word chunks and word-marker pairs are statistically learned and how they modulate reading performance and language processing. Designed self-paced reading and statistical-learning experiments and presented findings at the International Conference on Interdisciplinary Advances in Statistical Learning (ISLA 2022, 2024).

Multilingual FrameNet Project

International Computer Science Institute (ICSI), UC Berkeley. PI: Terry Regier, Collin Baker. Aug. 2019 – Dec. 2019.

Studied the structure of English lexical databases and ran model trainings on nine types of frame-to-frame semantic relations. Evaluated poorly predicted relations and constructed new relations from lexical units to improve model coverage; refined labeling of semi-automatic semantic-role and universal semantic frames to improve multilingual alignment.

Rule Generalization at Different Boundary Levels

Experimental Phonology Group, UC Berkeley. Advisor: Jesse Zymet. Aug. 2019 – May 2020.

Empirical and modeling work on how phonological rules generalize across morphological and prosodic boundaries, contributing to a broader theoretical project on the locality of phonological computation.

Acoustic Analyses of Iquito

Indigenous Language Revitalization Group, UC Berkeley. Advisor: Christine Beier. Aug. 2018 – May 2021.

Conducted acoustic analyses (Praat, ELAN) and social-background research on Iquito, an endangered Zaparoan language of Peru. Presented an Optimality-Theoretic analysis of verbal tone interactions at the 28th Manchester Phonology Meeting (2021) and a dispersion-theoretic model of tone systems at the 56th Linguistics Colloquium (2020).

Locomotion and Early Language Acquisition

Infant Study Center, UC Berkeley. PI: Joseph Campos. Jan. 2018 – Dec. 2019.

Studied the relationship between infant locomotion (walking onset), social-communicative abilities, and vocabulary growth in Chinese-American infants. Co-presented two posters at the 22nd Biennial International Conference of Infant Studies (2020).

Cost of Phonetic Cues in Mandarin–English Code-Switching

Berkeley PhonLab, UC Berkeley. PIs: Susan Lin, Alice Shen. Aug. 2017 – Dec. 2017.

Acoustic and behavioral investigation of the phonetic cost of code-switching between Mandarin and English, contributing to broader work on bilingual phonetic accommodation.