I build systems for language and creative work. These studies examine what changes when people use them; my next question concerns creative intentions that change during collaboration.
01 / SELECTED RESEARCH
Three studies. Different responsibilities.
Figure 5, published Sinolingo paper · task 1 / task 2 comparison
Figure 5, published Sinolingo paper · task 1 / task 2 comparison
First authorPublished
Can a translation platform support learning as well as output?
First author; contributed to the translation platform and the study reported in the paper.
Method
30 translation undergraduates completed three stages: conventional tools, Sinolingo, then a new task after one month of autonomous use. Outputs were compared with BLEU, chrF++ and BERTScore.
Finding
The paper reports BLEU 21.4 → 28.9, chrF++ 48.7 → 56.2 and BERTScore 0.812 → 0.864 from the first to second task.
Paper details & limitations
The Construction of the “Sinolingo” Platform and a Study on Its Effectiveness in Autonomous Translation Learning
Mingkai Li, Qiaoxi Huang, Jinwen Hu, Renjia Chen, Menglian Liu & Xingyu Pu
ETMIS 2025 · Atlantis Press · 2026
A small, single-institution study with sequential tasks and a short intervention. These are task-level findings, not proof of lasting learning for all students.
Published paper, Figure 1 · the corpus management interface
Published paper, Figure 1 · the corpus management interface
Second authorPublished · EI Compendex
What makes a bilingual corpus useful in translation practice?
Second author and project team member; contributed to corpus-platform development.
Method
News collection, translation, automatic alignment and human review produced 11,000 sentence pairs. A six-week evaluation compared 32 Vietnamese-major students, with 16 in each group.
Finding
Reported alignment accuracy: 96.5%. Mean post-test translation scores were 89.1 with corpus-supported practice and 82.4 in the comparison group.
Paper details & limitations
Construction and Application of a Vietnamese-Chinese Parallel Corpus for Political News
Xiaoxin Chen, Mingkai Li, Wenjing Liu, Xiaoyun Liu, Ke Wang & Menglian Liu
DEIT 2026 · ACM ICPS
One cohort, a short intervention and political-news texts. The result does not establish gains across other domains or long-term retention.
Current Echoo public website, October 2026 · a later product version, not the experiment interface
Current Echoo public website, October 2026 · a later product version, not the experiment interface
Fourth authorAccepted with minor revisions · editorial production pending
When do visual cues help an interpreter?
Fourth author; developed the real-time subtitles, trigger-based highlighting and supporting audio/transcription services.
Method
19 postgraduate trainee interpreters worked with and without highlighting in both Chinese–English directions. The study assessed accuracy, fluency, cognitive load and user acceptance.
Finding
The manuscript reports higher accuracy and lower perceived load. Trigger-specific gains were significant for proper nouns and terms, but not for lists or numbers.
Paper details & limitations
Human-centred, augmented simultaneous interpreting: supporting trainees with trigger-based highlighting
W. Su, Z. Chen, B. Zhu, M. Li & D. Li
Linguistica Antverpiensia, New Series – Themes in Translation Studies
Prepared trigger pairs, controlled speeches and a small trainee sample. My contribution was the supporting software, not leadership of the full experimental design.
Preference-guided generation can refine what a person already likes. My proposed study asks what happens when later feedback challenges those earlier choices: should the agent clarify, explore alternatives, or keep refining?
The mechanism to test
Separate fixed product facts, confirmed creative decisions and provisional preferences. Compare a preference model that can reconsider earlier choices with cumulative and recency-weighted baselines; evaluate action selection separately.
How I would evaluate it
One session, one product visual. Compare the same models and budget across a conversational agent, fixed workflow, preference-guided baseline and adaptive method. Evaluate novelty, usefulness, factual fidelity, control and ownership; use ablations to isolate the two mechanisms.
This is a planned study. Interviews and a pilot would refine the protocol and sample size; the interactive sketch below has not been evaluated with participants.
Research context & scope
The proposal builds on preference adaptation and adaptive elicitation, while focusing on reconsidering earlier choices and deciding when to reopen exploration. The initial scope excludes video, multi-agent architectures and long-term personalisation.
