Interactive Lexical Graphs for Supporting Vocabulary Growth in Language Learning

Vocabulary acquisition research consistently shows that contextually embedded word learning outper-forms decontextualised memorisation, yet most digital vocabulary tools still present words in isolated pairs-a pedagogically outdated design that discards the relational information most predictive of long-term retention. This paper builds interactive lexical graphs over a corpus drawn from 1,800 language learners' vocabulary logs (48,000 sessions), constructing a cooccurrence and se-mantic-relation graph of 48,000 nodes and 284,000 edges, and trains a Graph Neural Network (GNN) on the resulting structure to predict vocabulary-growth trajectories and recommend contextually optimal next words. Against held-out vocabulary-assessment scores, the GNN achieves Pearson$\mathbf{r}=\mathbf{0. 8 6 4}$, outperforming a standard LSTM vocabulary model ($\mathbf{r}=\mathbf{0. 7 6 4}$; +13.1%), Word2Vec embeddings (0.684; +26.3%), and a flashcard baseline (0.524; +64.9%). Deployed in a live vocabulary-learning application, retention at seven days rose from 52.4% to 78.4% (+49.6%) and active productive-use rate from 42.4% to 68.4% (+61.3%), relative to flashcard-system users. Semantic-recommendation precision reached 84.4% versus 62.4% for random co-occurrence selection (+35.3%). Cold-start accuracy degrades by 16.2% for learners with fewer than ten sessions, and low-frequency words show a 9.3% drop, each with a deployed operational mitigation.

Hali tarjima qilinmagan