Every language learner who has asked a chatbot to explain a word knows the frustration: the answer is helpful, the explanation is elegant, and then the conversation scrolls away, taking the hard-won linguistic insight with it. A team of researchers from Taiwan now argues that this ephemerality is not just an annoyance but a structural flaw in how large language models are being used for education. Their answer is Morphowiki, an open-source framework described in the journal SoftwareX, which converts fleeting AI dialogue into a permanent, version-controlled, morpheme-centered vocabulary network that learners can grow, audit, and review for years.
The system targets a skill that decades of second-language acquisition research have identified as foundational: morphological awareness, the ability to slice unfamiliar words into their meaningful building blocks. A learner who recognizes that “international” decomposes into the prefix inter-, the root nat, and the suffixes -ion and -al can decode whole families of words instead of memorizing them one by one. Large language models, the authors note, are remarkably good at performing this decomposition in real time, yet the conversational interfaces through which we access them leave the resulting knowledge trapped in unstructured chat logs, fragmented and impossible to consolidate systematically.
Morphowiki’s architecture is deliberately simple and rests on three layers. A generation layer orchestrates seven Agent Skills, small callable command modules that direct the language model through specific tasks. A data layer stores everything as plain-text Markdown files organized into directories for words, prefixes, roots, and suffixes. A review layer runs a lightweight local Node.js server with a single-page web application and a JSON datastore that tracks spaced-repetition scheduling. Because the entire knowledge base lives in Markdown and JSON rather than a proprietary database, it can be read by humans, parsed by programs, diffed by version-control systems, and shared freely, in sharp contrast to the closed data silos of commercial study platforms.
The workflow follows an “initialize, ingest, expand, review” cycle. After a one-command setup that builds the directory structure and installs dependencies, the learner feeds in an unfamiliar word or an entire sentence. The system tokenizes the input, filters out stopwords, both a default list of roughly 150 English function words and a personalized exclusion list the learner can extend, and then applies what the authors call an architecturally bounded recursive decomposition. Given “international,” the pipeline isolates the root nation (itself nat plus -ion), expands to national, and resolves the target word with the prefix inter-, creating a separate Markdown page for each layer and wiring every page bidirectionally to its constituent morphemes. A visited-set mechanism prevents the recursion from looping endlessly.
Once a word is archived, the /flashcard command enrolls it into the review queue after a strict three-part check: the word’s Markdown page must exist, the word must not be on the stopword list, and no duplicate entry may exist in the JSON datastore. Scheduling follows the classic SuperMemo-2 algorithm, initialized with an easiness factor of 2.5 and a zero interval. When the learner runs the review command, a local server identifies due cards, parses the relevant Markdown on the fly, and presents a four-button evaluation, Again, Hard, Good, or Easy. The authors condensed the traditional six-grade SM-2 scale with a mathematical justification: all failed retrievals reset the interval and repetition counters without touching the easiness factor, so they can safely be merged into a single “Again” button.
Two further commands push the system beyond passive flashcards. The /new-word command scans the learner’s accumulated morpheme pool and synthesizes derivatives that are not yet recorded, for example proposing “nationalize” once the root nat has been mastered, then automatically running the ingestion and linking pipeline. Because language models can hallucinate plausible-sounding pseudowords, the skill specification embeds a validation step requiring the model to confirm that each candidate is an attested modern English word. A /lint command then audits the knowledge base in two modes: a deterministic, token-free structural check of link symmetry and schema completeness, and a model-assisted semantic inspection of individual entries. Crucially, lint never edits anything on its own; it only raises flags, and every state change requires explicit human authorization.
What elevates the paper beyond a software description is its quantitative evaluation of the decomposition step. The authors assembled 950 test items, including a main set of 600 words stratified by vocabulary level and morpheme count, drawn from the intersection of a Taiwanese senior-high-school word list and the SIGMORPHON 2022 morpheme segmentation shared task. Against the unsupervised baseline BPE, which scored an F1 of 29.13, and Morfessor 2.0, which reached 47.26, Morphowiki paired with Claude Opus 5 achieved an F1 of 83.06 and an exact-match rate of 74.83 percent. On a subset of 200 words containing four or more morphemes, the baselines managed essentially nothing, 0.00 and 3.00 percent exact match, while Morphowiki still resolved 46.50 percent correctly.
The most intellectually interesting result comes from a two-step ablation that separates the contribution of the underlying model from that of the Agent Skills scaffolding. With the model held fixed, removing the skill file and then the repository’s project instructions revealed that the two components do very different work. The project instructions, which describe the project in etymological terms but prescribe no segmentation rules, actually degraded every measure on the main set and nearly tripled over-segmentation of simple words, from 6.7 to 23.3 percent. The skill file, by contrast, added more than 52 points of exact-match accuracy on words recorded in canonical form and corrected 17 of the over-segmentation errors the instructions induced. The authors conclude that the skill does not teach the model how to segment strings, a capability it already possesses, but rather imposes canonical discipline: knowing when to restore a base to its dictionary form, when to record an assimilated prefix canonically, and where recursion should stop.
The team is candid about the system’s limits. Decomposition quality remains tethered to the reliability of the underlying model, and the deployed configuration still over-segments monomorphemic words at 12.0 percent, a trade-off traced to the etymological priming of the project instructions rather than the skill itself. The reference standard used for scoring is itself noisy, with coverage of the most elementary vocabulary at only 43.1 percent, so the reported accuracies likely underestimate true performance. Cross-model comparison covered only two models from a single provider, the pipeline is calibrated exclusively for English derivational morphology, and cost is nontrivial: roughly 9 to 23 US dollars per thousand words processed, which positions the framework for curated vocabulary construction rather than bulk corpus processing. Pedagogical efficacy, finally, has not yet been tested in classrooms.
Even with those caveats, Morphowiki speaks to a tension now rippling through education research. Recent studies on cognitive offloading and “metacognitive laziness” suggest that instantaneous AI answers can undermine the deep semantic engagement that durable memory requires, a concern rooted in the classic levels-of-processing work of Craik and Lockhart. By forcing learners through active retrieval, spaced review, and human-approved curation, Morphowiki is designed to make the AI a builder of infrastructure rather than a dispenser of answers. The authors frame it as a reproducible testbed on which researchers in computer-assisted language learning can systematically compare models, prompting strategies, and scheduling algorithms on a standardized, transparent data format. If the broader lesson holds, the future of AI-assisted learning may depend less on how cleverly a chatbot explains a word, and more on whether the explanation leaves behind something the learner can keep.
Subject of Research: An open-source Agent Skills framework that uses large language models to decompose English vocabulary into morphemes, build bidirectionally linked knowledge graphs, and schedule spaced-repetition review.
Article Title: Morphowiki: reusable agent-skill infrastructure for morphology-centered vocabulary knowledge construction
Article References: Chen, Y.-T., Yao, K.-C., Chang, S.-H., & Lin, Y.-C. (2026). Morphowiki: reusable agent-skill infrastructure for morphology-centered vocabulary knowledge construction. SoftwareX, 36, Article 103114. https://doi.org/10.1016/j.softx.2026.103114
Image Credits: AI Generated
DOI: 10.1016/j.softx.2026.103114
Keywords: Morphowiki, large language models, morphology, vocabulary acquisition, spaced repetition, Agent Skills, computer-assisted language learning, open-source software, second language acquisition, cognitive offloading, Markdown knowledge graph, SM-2 algorithm
News Source: Denise Maddox. (October 12, 2026). Open-Source AI Framework Turns Chatbot Vocabulary Lessons Into a Lasting Knowledge Graph. Scienmag.
