DeepSeek’s R1 model handles out-of-vocabulary (OOV) words primarily through subword tokenization, a technique that breaks down unknown words into smaller, recognizable units. Instead of relying solely on a fixed vocabulary, the model uses algorithms like Byte-Pair Encoding (BPE) or similar methods to split words into subword fragments. For example, a word like “unsplash” might be decomposed into subwords like "un", "spl", and “ash” if those components exist in the training data. This approach allows the model to process words it hasn’t explicitly seen before by leveraging familiar subword patterns, reducing the impact of OOV scenarios.
The model’s tokenizer is trained to identify statistically frequent subword units during preprocessing. When encountering an OOV word during inference, the tokenizer applies the same rules to split the word into known subwords. For instance, technical terms like “TransformerXL” could be split into "Transform", "er", and "XL", assuming those subwords were part of the training corpus. This method ensures that even novel terms are represented through combinations of learned embeddings. Additionally, the model’s architecture—typically a transformer-based design—processes these subword sequences in context, allowing it to infer meaning based on surrounding tokens. This is critical for handling domain-specific jargon or newly coined terms that weren’t present in the training data.
For cases where subword decomposition isn’t sufficient (e.g., entirely novel character combinations), the model might employ fallback strategies. One common approach is to map rare or unsegmentable tokens to a special “unknown” token (e.g., <UNK>), though this is minimized by the subword approach. The R1 model likely supplements this with contextual attention mechanisms: even if a word’s subwords are unfamiliar, the transformer layers can use positional embeddings and attention weights to approximate meaning based on syntactic and semantic patterns in the sentence. For example, in the phrase “The quantum flux capacitor activated,” even if “flux capacitor” is split into subwords, the model could infer it’s a technical device based on “quantum” and “activated.” This combination of subword tokenization and contextual analysis balances robustness with computational efficiency.