Byte Pair Encoding
Byte Pair Encoding (BPE) is a subword tokenization algorithm that iteratively merges the most frequent pairs of characters or character sequences. BPE is widely used in modern language models to handle rare words and multilingual text.
Understanding Byte Pair Encoding
Byte Pair Encoding is a subword tokenization algorithm that has become the standard method for converting raw text into the token sequences that large language models process. Starting from individual characters, BPE iteratively identifies and merges the most frequently co-occurring pair of tokens in the training corpus, building a vocabulary of subword units that balance between full words and individual characters. Common words like "the" remain intact, while rare or compound words are split into meaningful subword pieces, allowing the model to handle previously unseen words through composition. For example, "unhappiness" might be tokenized as "un" + "happiness" or "un" + "happi" + "ness" depending on the vocabulary. BPE is used by GPT models, while BERT uses a related algorithm called WordPiece. The vocabulary size is a key hyperparameter that trades off sequence length against vocabulary coverage, with typical sizes ranging from 30,000 to 100,000 tokens for modern transformer models.
Category
Natural Language Processing
Is AI recommending your brand?
Find out if ChatGPT, Perplexity, and Gemini mention you when people search your industry.
Check your brand — $9Related Natural Language Processing Terms
Abstractive Summarization
Abstractive summarization generates new text that captures the key points of a longer document, rather than simply extracting existing sentences. It requires deep language understanding and generation capabilities.
Beam Search
Beam search is a decoding algorithm that explores multiple candidate sequences simultaneously, keeping only the top-k most promising at each step. It balances between greedy decoding and exhaustive search in text generation.
BERT
BERT (Bidirectional Encoder Representations from Transformers) is a language model developed by Google that reads text in both directions simultaneously. BERT revolutionized NLP by enabling deep bidirectional pre-training for language understanding tasks.
Bigram
A bigram is a contiguous sequence of two items (typically words or characters) from a given text. Bigram models estimate the probability of a word based on the immediately preceding word.
Corpus
A corpus is a large, structured collection of text documents used for training and evaluating natural language processing models. The quality and diversity of a training corpus significantly impacts model performance.
Extractive Summarization
Extractive summarization selects and combines the most important sentences directly from a source document to create a summary. It preserves the original wording but may lack the coherence of abstractive approaches.
Grounding
Grounding in AI refers to connecting a model's language understanding to real-world knowledge, data, or sensory experience. Grounded AI systems produce more factual and contextually relevant outputs.
Language Model
A language model is an AI system that learns the probability distribution of sequences of words in a language. Modern language models like GPT and Claude can generate text, answer questions, and perform complex reasoning.