This flagship project underpins the ALL’s initiative to develop an integrated AI resource hub and a suite of chatbots to address writing challenges among self-financing tertiary institutions in Hong Kong, where many learners struggle with fundamental aspects of writing such as cohesion, formality, register, and genre-specific conventions in both Chinese and English. Unlike general-purpose AI tools, which tend to offer surface-level corrections, the solutions offered by ALL emphasize pedagogical, genre-specific support for writing in both languages. 

The LinguaLink ALLiance (博文智匯) is an online AI tooklit curating free resources (e.g., Effidit, Shenyandayi深言達意) for supporting writing development and multi-modal learning. Our team rigorously evaluates tools collected for the LinguaLink ALLiance specifically from the perspective of Hong Kong academics and students, assessing each tool for script compatibility, genre accuracy, pedagogical value, and contextual suitability for the Hong Kong linguistic environment. Our evaluations are shared through mediums such as user case reports and review videos, which are made freely available on the platform for the benefit of the wider Hong Kong education community.

The project’s other core deliverable, the adaptive learning chatbot Aybo (阿博), provides scaffolded feedback to support revision and skill enhancement without requiring prior genre knowledge or manual input of assignment criteria, drawing on course-specific guidance to tailor its support to individual students. Crucially, Aybo offers this individualized, genre-aware feedback without directly modifying students' text, encouraging autonomous revision and skill consolidation. Through iterative interactions with Aybo, students receive scaffolded recommendations targeting coherence, lexical choice, and structural appropriateness within specific genres to improve their own writing. In the 2025/2026 academic year, Aybo has been implemented across six university-wide Chinese and English language courses, reaching over 1,800 undergraduate students. Our team has observed measurable improvements in writing quality from using Aybo, while also uncovering key lessons about the capacities and constraints for implementing custom AI tools in large-scale writing instruction.

Recognising ongoing student challenges in writing formal Chinese correspondence—particularly conventions regarding letter formats, honorific and salutatory expressions (敬語/祝頌語), formal register, and opening and closing forms of address (上款/下款)—this project develops adaptive chatbots to support formal Chinese letter writing. General-purpose AI tools (e.g., ChatGPT, ERNIE 文心一言) currently fall short in addressing these challenges: they mix simplified and traditional scripts, insert colloquial language, and fail to reliably generate appropriate forms of salutations and closings. Our 2023–2024 preliminary study confirms such limitations, with generic LLMs producing only stock expressions, while struggling with contextually correct salutatory forms.

To address these issues, this team has developed the Aybo Chinese Invitation Letter Writing Assistant (阿博邀請函寫作助手) and the Aybo Chinese Cover/Job-Application Letter Writing Assistant (阿博求職信寫作助手), deployed on the LinguaLink ALLiance (博文智匯) platform via Microsoft Azure. Drawing from course materials and authoritative writing guidelines from the Hong Kong Education Bureau, these writing assistants employ an interactive guided-writing interface grounded in the principle of “no ghost-writing”. This operating model engages students incrementally, providing real-time feedback and worked examples to foster mastery of letter structure and writing etiquette, without substituting for instructor input or independent learning.

Evidence of Impact

The writing assistants have been evaluated through quasi-experimental classroom intervention activities with students enrolled in the First Year Chinese (大一國文) course. In the first semester of the 2025/26 academic year, the team conducted a pre-/post-test study with the Invitation Letter Writing Assistant, comparing 266 students across three treatment groups. Participants in the treatment group with the greatest exposure to Aybo, including extended out-of-class use over four weeks, demonstrated significant gains in knowledge of letter structure and recognition of forms of address. Moreover, a single 30-minute session with Aybo boosted overall vocabulary recognition of salutatory terms, while extended use drove more precise discrimination of contextually correct and incorrect options. These findings highlight that in-class use plus ongoing practice with AI-powered writing assistants can lead to superior learning outcomes. A parallel study with the Cover/Job-Application Letter Writing Assistant took place during the second semester of the 2025/26 academic year.

The team has developed and evaluated the SYU Cantonese Segmentation Chatbot (SYU-CSC, 廣東話分詞系統), an AI tool addressing the core challenge of word segmentation in Cantonese. Correctly recognizing word boundaries in Cantonese is essential for downstream language tasks such as speech recognition, translation, and linguistic analysis. Complicating the segmentation problem is the “modern diglossia” phenomenon in Hong Kong Cantonese, where spoken Cantonese and Standard Written Chinese (SWC) intermingle in vocabulary, registers, and even character forms. 

Conventional segmentation tools (e.g., PyCantonese, Cybercan, Cantoseg) rely on fixed dictionaries, being unable to keep pace with evolving slang and colloquial language, as well as requiring advanced technical skills to operate. SYU-CSC instead leverages the contextual learning capabilities of large language models (LLMs) to segment text, assign parts of speech, and classify lexical items according to registers (exclusively Cantonese, exclusively SWC, or common to both), providing English translations for verifying meaning accuracy. Requiring no programming expertise to operate, SYU-CSC delivers comprehensive, consistently formatted segmentation outputs from direct text input, responding with natural language personas in a single integrated workflow. The chatbot is now accessible through platforms such as Chipp.ai, Google Gemini, and POE.

Evidence of Impact

In 2025, SYU-CSC was benchmarked against four leading platforms (Doubao, ERNIE 文心一言, Gemini, and GPT-4o), using 45 sentences drawn from three landmark works of Hong Kong literature spanning 1950–2020.

• Superior accuracy: SYU-CSC achieved the greatest accuracy in both segmentation (87.92%) and part-of-speech tagging (97.12%). Notably, it significantly outperformed every comparison tool in lexical register classification accuracy (78.87%). A follow-up empirical study found that SYU-CSC outperformed general-purpose and Chinese-optimized LLMs by 15–18% in lexical classification accuracy, especially excelling at handling inter-dialectal homographs. 
• User experience: SYU-CSC needs only direct input of the text requiring segmentation to return accurate and consistently formatted results where other tools require repeated prompting, achieving a no-prompt workflow.

The project team is currently working on segmenting, tagging, and classifying the words collected in the Hong Kong Cantonese Corpus (HKCanCor, 2015) by Luke and Wong to further verify the performance of SYU-CSC on spoken Cantonese corpus data at scale.

The team has developed and empirically validated Generative Adversarial Authentication (GAA), a low-cost method for instructors to detect improper AI use in student assignments. GAA is not opposed to legitimate AI-assisted learning, but rather targets “AI abuse”, where students submit AI-generated work blindly without meaningful effort to understand or revise the text beforehand. The urgency of tackling AI abuse arises because advanced AI writing tools can now produce high-quality, even creative essays indistinguishable from authentic student work. Moreover, current automated detectors (e.g., VeriGuide, Turnitin, ZeroGPT) are unreliable, flagging texts inconsistently and struggling particularly with “AI+human” hybrid writings. This leaves teachers in a bind: with no trustworthy proof of AI use, they risk either falsely accusing students or overlooking real violations of academic integrity.

GAA offers a fundamentally different “fight-AI-with-AI” strategy inspired by generative adversarial networks (GANs). Instead of analyzing the text itself, GAA assesses authorship, using AI to generate content-specific multiple-choice questions about students’ own submissions. These questions are given to students as a surprise quiz; a low error rate suggests that a student genuinely understands the work they have submitted, whereas a high error rate is evidence of superficial authorship. By probing cognitive retention among authors rather than writing features in texts, GAA provides an effective instrument for detecting over-reliance on AI, even against advanced paraphrasing of AI-generated text.

Evidence of Impact

A double-blind study (2023–2025) with 96 Hong Kong Shue Yan University undergraduates compared GAA to two traditional AI detectors, VeriGuide and ZeroGPT.

• GAA alone reached 90.91% accuracy, on par with VeriGuide and ahead of ZeroGPT (84.85%).
• GAA combined with VeriGuide reached 93.94% accuracy, outperforming any single detection method, using a GAA error threshold of 40% and a VeriGuide AI index threshold of 40–60%.
• Practical and efficient: on average, students completed a 10-question multiple-choice quiz in ≤3 minutes. Teachers require no prompt-engineering skills to generate the quizzes, and teaching assistants can quickly grade quiz submissions.
• Flexible deployment: GAA can be implemented via randomized sampling, instructor activation, or automated tool-based triggering.
• GAA detects “free riders” in group work: a bonus finding—members unfamiliar with the text content display more errors on the quiz, exposing their limit contributions to the group submission.

Ongoing enhancements to GAA include improving objective question-generation strategies and generating “counter-test” items on quizzes to counteract memorized responses.

The team has developed and evaluated the SYU Traditional Chinese Writing Assistant (SYU-TC, 繁體中文寫作助手), an AI-powered error-detection and proofreading tool designed for Traditional Chinese users in regions such as Hong Kong, Macau, and Taiwan. As most existing Chinese spell checkers and writing assistants are trained predominantly on Simplified Chinese text data, they underperform in recognizing Traditional Chinese errors and handling mixed-script texts. SYU-TC addresses this technical gap, being built on a re-parameterized version of DeepSeek-V3 and delivered via a chatbot interface. SYU-TC detects five types of writing errors: (1) character/word-level mistakes (typos, homophone and look-alike confusions, punctuation), (2) grammatical errors (redundant, missing, or misordered components), (3) semantic errors (poor word choice and collocation), (4) Traditional–Simplified normativity errors (wrong Traditional characters and Simplified–Traditional mixing), and (5) Cantonese-colloquialisms-in-writing (粵口語入文) — where spoken Cantonese vocabulary intrudes into formal written Chinese.

A defining feature of SYU-TC is its no-prompt (自動檢測) workflow for error detection: users just have to input the target text, and SYU-TC automatically detects errors and suggests corrections without requiring repeated prompting. Applying a uniform detection standard to every input, SYU-TC embodies ALL’s user-centered philosophy, shifting people from adapting to machines, to machines serving people.

Evidence of Impact

SYU-TC was benchmarked using a test corpus of 394 sentences (~35,500 characters). Means, standard deviations, and ANOVAs for error detection accuracy were calculated for SYU-TC and five leading AI tools (DeepSeek-R1, DeepSeek-V3, GPT-4o, ERNIE (文心一言), and Doubao.

• Greatest overall accuracy: SYU-TC achieved the greatest overall detection accuracy (92.40%) across all error categories (vs. 78.40% for GPT-4o) with the lowest variability.
• Mixed-script normativity errors: SYU-TC achieved 98.00% detection accuracy, far ahead of four LLMs (27–50%), and significantly outperforming GPT-4o (91.90%).
• Cantonese-colloquialisms-in-writing errors: SYU-TC achieved 86.40% detection accuracy, significantly outperforming all comparison tools.
• Grammatical and semantic errors: SYU-TC achieved 93.60% and 96.3% detection accuracy respectively and demonstrated the most consistent performance.

This benchmarking study highlights that task-specific models trained on high-quality data can surpass general-purpose large language models (LLMs) in specialized vertical domains, challenging the assumption that bigger models necessarily equal better performance.