Fangzhou Zhai

Research Associate - Script Knowledge at German Research Foundation (DFG) – Saarland University (UdS)

Rese arch focus: Investigate acquisition, formalization and application of common-sense knowledge scripts for NLP, including the construction of high-quality annotations, formalization of scripts, and applications to story generation and QA. Led development of a zero-shot script parser, designed metrics for cluster coherence and performance, and achieved results comparable to prior supervised models on unseen Script Events; contributed to MCScript cross-corpus generalization.

Research Associate - Script Knowledge at German Research Foundation (DFG) with Saarland University (UdS)

Investigated acquisition of common sense knowledge (scripts) for NLP; developed the first zero-shot script parser using domain-adapted XLNet embeddings and clustering to enable event/entity recognition in unseen everyday scenarios (e.g., taking a bath, grocery shopping). Designed novel metrics to evaluate cluster coherence and performance. Achieved results competitive with prior supervised models on InScript events and demonstrated cross-corpus generalization on MCScript.

Recipe Alignment at Computational Linguistics Group, Saarland University (UdS)

Designed and deployed a web-based annotation framework to align procedural steps across 110 recipes spanning 10 dishes; captured verb/predicate alignments and global procedural structure; monitored annotation process and produced the first aligned recipe data set for downstream knowledge formalization (EM NLP 2021).

Story Generation with Rich Details at German Research Foundation (DFG) – Saarland University

Designed a dual-decoder seq2seq system (outline + detail) with MMI objective for informative narratives; developed the scheme and a web UI to evaluate quality of machine-generated stories by human participants; improved informativeness scores while maintaining coherence.

Script Parsing Supervisor at German Research Foundation (DFG) – Saarland University (UdS)

Built hierarchical BERT / XLNet + BiLSTM model capturing local patterns and global procedural structure; significantly outperformed prior SOTA on event detection; established first benchmarks for script participant identification and contributed to EMNLP/ACL-level work.

Research Lead Investigator - Optical Recognition at Spoken Language Systems, SIC

Led data collection and model development for optical recognition within the SL Systems group; oversaw project design and evaluation for recognition tasks.

Research Associate - Eye-Tracking & Gaze Detection at DFKI

Developed approaches to verify the plausibility of automatically calibrating eye-tracking devices; clustered gaze points within each time window around objects of interest to determine spherical coordinates of gaze.

Eye-Tracking & Gaze Detection Research Associate at DFKI

Developed an approach to verify plausibility of automatically calibrating eye-tracking devices; clustered gaze points within time windows by objects of interest to determine the spherical coordinates of gaze.

Research Associate - Probabilistic Model of Reasoning at University of Amsterdam (UvA)

Constructed a probabilistic proof system to explain human reasoning behaviors; parameterized cognitive preferences to be trained with EM. Demonstrated NP-complete problems admit large simple fragments and highlighted hierarchical increases in expressive power of dynamic logics.

Probabilistic Model of Reasoning Research Associate at University of Amsterdam

Constructed a probabilistic proof system to explain human reasoning behaviors; parameterized cognitive preferences to be trained with EM; demonstrated NP-complete problems admit large simple fragments with polynomial-time explainability; showed hierarchical increases in expressive power of dynamic logical extensions.

Intern, Research Internship - Recommendations Feed at Wang Xuan Institute of Computing Technology

Developed a blog feed classifier distinguishing organizational vs. personal blogs; implemented a logistic regression model with engineered features (named entities, proper names, typographic cues, stylistic cues); earned 2nd place in the 2010 TREC Blog Track, contributing to advances in document retrieval and large-scale sentiment analysis.