WordCraft: Scaffolding the Keyword Method for L2 Vocabulary Learning with Multimodal LLMs
Authors
Paper Title
WordCraft: Scaffolding the Keyword Method for L2 Vocabulary Learning with Multimodal LLMs
Publication Info
- Topic area: Enhancing second-language vocabulary learning through interactive tools and multimodal large language models (MLLMs).
- Keywords: Keyword method, L2 vocabulary learning, multimodal LLMs, cognitive scaffolding, generation effect, creativity support tools, human-computer interaction, memory retention, semantic associations, image formation.
Background and Problem
- Problem / challenge: L2 learners face difficulties in applying the keyword method due to challenges in keyword selection, association construction, and vivid mental imagery creation. Existing computational tools either reduce learner engagement or fail to provide adequate scaffolding for the generative process.
- Significance: Effective vocabulary retention is crucial for language learning, and the keyword method has proven benefits. However, its practical application is hindered by high cognitive demands, limiting its effectiveness.
- Motivation and related work: Prior research has validated the keyword method’s efficacy but highlighted its challenges. Computational enhancements have focused on either fully automating the process or augmenting outcomes, often at the cost of the generation effect. Current creativity support tools and MLLM integrations lack alignment with the iterative and non-linear demands of the keyword method.
Solution
- Proposed approach: WordCraft, an interactive tool powered by MLLMs, designed to scaffold the keyword method by guiding learners through keyword selection, association construction, and image formation.
- Novelty:
- Process-level scaffolding that balances cognitive load and preserves the generation effect.
- Integration of MLLMs for multimodal support, including semantic brainstorming, association mapping, and visual imagery.
- Structured workflow enabling iterative and adaptive learning across stages.
- Procedure and key techniques:
- Keyword Selection: Phonological segmentation, semantic brainstorming, and keyword exploration with MLLM suggestions.
- Association Construction: Structured association mapping with concept nodes, association links, and heuristic prompts.
- Image Formation: Visual-centric mental imagery canvas with recall paths and interactive mapping of keywords and meanings.
Results
- Concrete findings:
- WordCraft significantly improved long-term recall (delayed recall: M = 6.25 for self-generated cues vs. M = 4.20 for peer-generated cues).
- Higher ratings for generated cues in satisfaction, clarity, memorability, novelty, and relevance compared to baselines.
- Enhanced usability (SUS score: M = 6.06) and creativity support (CSI score: M = 6.19 for exploration).
- Advantage over baselines:
- Outperformed GPT-4o and flashcard interfaces in recall performance, cue quality, and user engagement.
- Preserved the generation effect, with self-generated cues showing stronger retention than peer-generated cues.
- Experiments / evaluation:
- Study 1 (N = 48): Compared WordCraft with GPT-4o and flashcard baselines in usability, learning outcomes, and recall performance.
- Study 2 (N = 20): Examined the generation effect by comparing self-generated and peer-generated cues.
- Limitations and future work:
- Limited to L1 Chinese–L2 English learners; cross-linguistic generalizability needs further exploration.
- Small vocabulary set; future studies should evaluate larger-scale, classroom-like scenarios.
- Constraints in supporting polysemous words and combining multiple mnemonic strategies.
Summary
WordCraft is an MLLM-powered tool designed to scaffold the keyword method for L2 vocabulary learning, addressing challenges in keyword selection, association construction, and image formation. It preserves the generation effect while enhancing usability, creativity, and long-term retention. Two user studies demonstrated its effectiveness compared to GPT-4o and flashcard baselines. While promising, future work should expand its linguistic applicability, scale, and flexibility to support diverse learning strategies and contexts.
Research Questions / Practical Problems
Question signals indexed for this paper.
- 71%
Co-designing Large Language Model Tools for Project-Based Learning with K12 Educators
CHI '25· Human-LLM Collaboration +2
- 63%
When Scaffolding Breaks: Investigating Student Interaction with LLM-Based Writing Support in Real-Time K-12 EFL Classrooms
CHI '26· Human-LLM Collaboration +3
- 63%
Do Teachers Dream of GenAI Widening Educational (In)equality? Envisioning the Future of K-12 GenAI Education from Global Teachers’ Perspectives
CHI '26· Human-LLM Collaboration +3
Based on Jaccard similarity of research subtopics & professions (≥60%)