WordCraft: Scaffolding the Keyword Method for L2 Vocabulary Learning with Multimodal LLMs

TL;DR

This study tackles durable L2 vocabulary retention for L1-Chinese learners by embedding the keyword method within a multimodal, AI-augmented workflow. It introduces WordCraft, a process-level creativity support tool that scaffolds keyword selection, association construction, and image formation to sustain active learner generation while leveraging MLLMs for guidance. Two user studies show WordCraft preserves the generation effect and yields higher usability, creativity support, and long-term recall compared with unstructured AI prompts and traditional flashcards, albeit with increased cognitive load and time requirements. The work highlights design principles for human–AI collaboration in L2 vocabulary learning, including maintaining learner agency, aligning visuals with mental imagery, and fostering a community of keyword-method practitioners. Overall, WordCraft demonstrates the potential of adaptive, multimodal scaffolding to enhance deep processing and durable memory in vocabulary education.

Abstract

Applying the keyword method for vocabulary memorization remains a significant challenge for L1 Chinese-L2 English learners. They frequently struggle to generate phonologically appropriate keywords, construct coherent associations, and create vivid mental imagery to aid long-term retention. Existing approaches, including fully automated keyword generation and outcome-oriented mnemonic aids, either compromise learner engagement or lack adequate process-oriented guidance. To address these limitations, we conducted a formative study with L1 Chinese-L2 English learners and educators (N=18), which revealed key difficulties and requirements in applying the keyword method to vocabulary learning. Building on these insights, we introduce WordCraft, a learner-centered interactive tool powered by Multimodal Large Language Models (MLLMs). WordCraft scaffolds the keyword method by guiding learners through keyword selection, association construction, and image formation, thereby enhancing the effectiveness of vocabulary memorization. Two user studies demonstrate that WordCraft not only preserves the generation effect but also achieves high levels of effectiveness and usability.

Figures (12)

• Figure 1: Procedure of the formative study. The study unfolded in four phases: 1) Introduction: Participants were introduced to the keyword method and familiarized with the system's operation. 2) Construction: Participants completed two keyword-construction tasks—first independently and then collaboratively with GPT-4o—while following a think-aloud protocol. 3) Interview: Semi-structured interviews were conducted to probe participants' cognitive processes and challenges. 4) Discussion: A focus group with learners, teachers, and researchers was held to gather insights on integrating MLLMs into the keyword method.
• Figure 2: WordCraft consists of four primary views: (A) Word Overview, which provides key information about the target word; (B) Keyword Selection, which supports keyword identification and fosters deeper comprehension; (C) Association Construction, which connects the selected keywords to the target meaning; and (D) Image Formation, which translates the keywords and meanings into visual elements to produce a final image.
• Figure 3: (A) Add Elementand (B) Associateoperations. (A1) Create an element using a click-and-drag action. (A2-A3) Click the plus button to open Image Content dialog, where users can select keywords and provide descriptions. (A4) The Recall Path the updates accordingly, and the element is visualized as a pie chart. (B1-B2) Once two elements are added, select them and annotate their relationship.
• Figure 4: Examples of the generated outputs using WordCraft and GPT-4o.
• Figure 5: Generated Clues Evaluation. (a) Completion Time with WordCraft is longer than with the GPT-4o baseline and Flashcard baseline, indicating deeper cognitive engagement. (b) Result Scoring shows that cues generated with WordCraft receive higher ratings across multiple dimensions compared to GPT-4o. Overall, all metrics demonstrate improvement over with the baseline.