LLMs handle out-of-vocabulary (OOV) words or tokens by leveraging their tokenization process, which ensures that even unfamiliar or rare inputs are represented in a way the model can understand. Here’s how it works:
1. Tokenization Techniques:
- Subword Tokenization (e.g., Byte Pair Encoding (BPE), WordPiece, Unigram):
- GPT or BERT use subword tokenization methods that break down words into smaller units (subwords). For example:
- The word “unfamiliar” might be split into
["un", "familiar"]. - A rare word like “autodidactism” might be split into
["auto", "did", "act", "ism"]. - This ensures that even if a word isn’t in the model’s vocabulary, its components are, allowing the model to process it effectively.
- Byte-Level Tokenization:
- Considers each byte (0-255) as a potential token.
- Models tokenize inputs at the byte level, allowing them to handle any text input, including misspellings, rare words, or text in various languages.
- For instance, “🤖AI🌍” would be tokenized into individual byte-level tokens representing emojis and characters.
2. Embedding and Context Understanding:
- Each token (subword or byte) is mapped to an embedding in a high-dimensional space.
- Even if the model encounters a novel word or token combination, the embeddings of its components allow it to infer meaning based on contextual patterns learned during training.
3. Handling Entirely Novel Inputs:
- Misspelled or Noisy Inputs:
- These are broken into tokens the model knows, which might still capture some semantic information. For example, “autdoidactism” could be tokenized similarly to “autodidactism.”
- Code or Specialized Notation:
- LLMs trained on diverse datasets (e.g., code snippets, technical writing) can handle uncommon tokens by identifying familiar patterns or structures.
4. Limitations:
- Loss of Semantic Specificity:
- If a rare or completely novel word is tokenized into many smaller units, its specific meaning might not be fully captured.
- Tokenization Overhead:
- Rare or OOV words may require multiple tokens, increasing computational cost.
- Dependence on Training Data:
- If the model hasn’t encountered similar patterns in training, its ability to infer meaning may be limited.
Example:
- Input: “ChatGPT is great for autodidactism!”
- Tokenization (BPE):
["Chat", "GPT", "is", "great", "for", "auto", "did", "act", "ism"] - Model uses embeddings for each token and context to process the sentence.
By tokenizing inputs into manageable units and leveraging context-aware embeddings, LLMs can handle OOV words effectively, even if not perfectly.