Discovering Lexical Gaps Using Embeddings from Multilingual LLMs

In three linesAutomated framework to detect lexical gaps (words absent in certain languages) using embeddings from multilingual LLMs. On Korean-English translation pairs, 4000 embedding spaces show gap words have weaker cross-lingual semantic alignment. Logistic classifiers achieve AUC 0.81–0.76 and retrieve 18/19 and 26/27 gap words.

Summary generated by Claude — human-verified