out-of-vocabulary (oov)

OOV (Out-of-Vocabulary) refers to words not in the models vocabulary, historically a major NLP challenge largely solved by subword tokenization. **Traditional problem**: Fixed word vocabularies could not handle unseen words, required UNK (unknown) token replacement, lost information. **With subword tokenization**: Words split into known subwords, virtually no true OOV. Cryptocurrency becomes crypto + curr + ency. **When OOV still occurs**: Character-level models with limited character set, very unusual Unicode, corrupted text. **Handling strategies**: **Traditional**: UNK replacement, spelling correction, stemming. **Modern**: Subword fallback to characters/bytes, byte-level tokenization guarantees no OOV. **Rare token issues**: While not technically OOV, rare subwords have poor embeddings due to limited training. **Code and technical text**: May contain identifiers and tokens underrepresented in training. **Evaluation consideration**: OOV rate used to measure vocabulary coverage on test sets. **Modern status**: Byte-level BPE and SentencePiece essentially eliminated OOV problem for text, shifting focus to rare token quality.