The same word written with a capital or lowercase letter can stump a powerful neural network. Why? Because it splits text into tokens and considers 'Cat' and 'cat' as different universes. This is vocabulary mismatch — one of the main issues in fast search.

In sparse search, the query and documents are turned into vectors where most numbers are zeros, and only keywords matter. It's as fast as a book's index, but sensitive to how the text is tokenized. New models like ModernBERT excel at heavy tasks, but for such search their vocabulary is too messy. Each word is split into dozens of variants: case, typos, rare forms. The neural network wastes energy on pointless cataloging instead of searching for meaning. It's a Tower of Babel where there is no common language.

The Vocabulary Transfer method acts like a translator. First, choose a simple intermediary vocabulary — like that of the old BERT model without case sensitivity, where 'Token' and 'token' are the same token. Then build a semantic bridge. For common tokens, copy the existing representations; for new ones, assemble them as the average of the nearest 'anchors'. A special algorithm (sparsemax) keeps only the strongest connections, zeroing out noise. Then the model is briefly fine-tuned: just 500 steps with masking (about 0.2% of the original training). The final touch is activation calibration: tweak the settings so that the network's neurons don't fall into permanent silence.

Result: ModernBERT-VT achieves 52.4 points on the nDCG scale on BEIR tests — a gain of 4.7 points. The method revives hopeless models. Another giant, RoBERTa-large, initially scored a meager 1.4 points, but after transfer it shoots up to 51.3. Interestingly, without activation calibration, everything drops to 48.8. So, overcoming the confusion of word forms, the neural network finally speaks the language of precise and fast search. And we find what we need faster.