Main Content

encodeTokens

Convert tokens to token codes

Since R2023b

    Description

    [tokenCodes,segments] = encodeTokens(tokenizer,tokens) encodes tokens using the specified tokenizer and returns the token codes and segments. This syntax automatically adds special tokens to the input.

    example

    [tokenCodes,segments] = encodeTokens(tokenizer,tokens1,tokens2) encodes the sentence pair tokens1,tokens2. This syntax automatically adds special tokens to the input.

    [tokenCodes,segments,idx] = encodeTokens(___) also returns the mapping between the input and the encoded output.

    ___ = encodeTokens(___,AddSpecialTokens=tf) specifies whether to add special tokens to the input.

    Examples

    collapse all

    Load a pretrained BERT-Base neural network and corresponding tokenizer using the bert function.

    [net,tokenizer] = bert;

    View the tokenizer.

    tokenizer
    tokenizer = 
      bertTokenizer with properties:
    
            IgnoreCase: 1
          StripAccents: 1
          PaddingToken: "[PAD]"
           PaddingCode: 1
            StartToken: "[CLS]"
             StartCode: 102
          UnknownToken: "[UNK]"
           UnknownCode: 101
        SeparatorToken: "[SEP]"
         SeparatorCode: 103
           ContextSize: 512
    
    

    Encode the tokens "Bidirectional", "Encoder", "Representations", "from", and "Transformers" using the encodeTokens function.

    tokens = ["Bidirectional" "Encoder" "Representations" "from" "Transformers"];
    [tokenCodes,segments] = encodeTokens(tokenizer,tokens);

    View the token codes.

    tokenCodes
    tokenCodes = 1x1 cell array
        {[102 7227 7443 7543 2390 4373 16045 2100 15067 2014 19082 103]}
    
    

    View the segments.

    segments
    segments = 1x1 cell array
        {[1 1 1 1 1 1 1 1 1 1 1 1]}
    
    

    Input Arguments

    collapse all

    Tokenizer, specified as a bertTokenizer or bpeTokenizer object.

    Input tokens, specified as a tokenizedDocument array, string array, or cell array of character vectors.

    Input sentence pairs, specified as tokenizedDocument arrays, string arrays, or cell arrays of character vectors.

    Flag to add padding, start, unknown, and separator tokens to input, specified as 1 (true) or 0 (false).

    Output Arguments

    collapse all

    Token codes, returned as a cell array of vectors of positive integers. The token codes index into the tokenizer vocabulary.

    Data Types: cell

    Segment indices, returned as a cell array of vectors of positive integers.

    The segments indicate which input token corresponds to which input. For each element s of the cell array, the value s(i) indicates which input corresponds to tokenCodes(i). If you specify a single string as input, then each element of segments is an array of ones.

    Data Types: cell

    Mapping between the token codes and inputs, returned as a vector of positive integers.

    The value idx(i) indicates which input token corresponds to tokenCodes(i). If tf is 1 (true) and tokenCodes(i) is the padding, start, unknown, or separator code of the tokenizer, then idx(i) is NaN.

    Data Types: double

    Algorithms

    collapse all

    WordPiece Tokenization

    The WordPiece tokenization algorithm [2] splits words into subword units and maps common sequences of characters and subwords to a single integer. During tokenization, the algorithm replaces out-of-vocabulary (OOV) words with subword counterparts, which allows models to handle unseen words more effectively. This process creates a set of subword tokens that can better represent common and rare words.

    These steps outline how to create a WordPiece tokenizer:

    1. Initialize vocabulary — Create an initial vocabulary of the unique characters in the data.

    2. Count token frequencies — Iterate through the training data and count the frequencies of each token in the vocabulary.

    3. Merge most frequent pairs — Identify the most frequent pair of tokens in the vocabulary and merge them into a single token. Update the vocabulary accordingly.

    4. Repeat counting and merging — Repeat the counting and merging steps until the vocabulary reaches a predefined size or until tokens can no longer merge.

    These steps outline how a WordPiece tokenizer tokenizes new text:

    1. Split text — Split text into individual words.

    2. Identify OOV words — Identify any OOV words that are not present in the pretrained vocabulary.

    3. Replace OOV words — Replace the OOV words with their subword counterparts from the vocabulary. For example, by iteratively checking that OOV tokens start with vocabulary tokens.

    Byte Pair Encoding

    Byte pair encoding (BPE) is a tokenization algorithm that allows transformer networks to handle a wide range of vocabulary without assigning individual tokens for every possible word. During tokenization, the algorithm replaces out-of-vocabulary (OOV) words with subword counterparts, which allows models to handle unseen words more effectively. This process creates a set of subword tokens that can better represent common and rare words.

    These steps outline the algorithm for training a BPE tokenizer:

    • Start with a corpus of text. For example, a corpus that includes phrases like "use byte pair encoding to tokenize text". Split the text data into words using a specified pretokenization algorithm.

    • Initialize a vocabulary of bytes. For example, start with a vocabulary of ["a" "b" "c" ... "z"]. For non-ASCII characters, like emojis that consist of multiple bytes, start with the byte values that comprise the character.

    • Encode each word in the text data as a sequence of bytes, and represent the words as sequences of integers that specify the indices of the tokens in the vocabulary. For example, represent the word "use" as [21 19 5]. When the encoding of a character is more than one byte, the resulting sequence of bytes can have more elements than the number of characters in the word.

    • Count the frequency of all adjacent pairs of bytes in the corpus. For example, among the words ["use" "byte" "pair" "encoding" "to" "tokenize" "text"], the token pairs ["t" "e"], ["e" "n"], and ["t" "o"] appear twice, and the remaining pairs appear once.

    • Identify the most frequent pair and add the corresponding merged token to the vocabulary. In the words represented as sequences of vocabulary indices, replace the corresponding pairs with the index of the new merged token in the vocabulary. Then, add this token pair to the merge list. For example, append the token pair ["t" "e"] to the merge list. Then, add the corresponding merged token "te" to the vocabulary so that it has the index 27. Finally, in the text data represented as vocabulary indices, replace the pairs of vocabulary indices [20 5] (which corresponds to ["t" "e"]) with the corresponding new vocabulary index:

      • The representation [2 25 20 5] for the word "byte" becomes [2 25 27].

      • The representation [20 5 24 20] for the word "text" becomes [27 24 20].

    • Repeat the frequency count and merge operations until you reach a specified number of iterations or vocabulary size. For example, repeating these steps several times leads to merging the pair ["b" "y"] to make the token "by", and then subsequently the pair ["by" "te"] to make the token "byte".

    These steps outline how a BPE tokenizer tokenizes new text:

    1. Pretokenization — Split text into individual words.

    2. Byte-encoding — Encode each word into sequences of bytes.

    3. Merge — By starting at the top of the merge list and progressing through it, iteratively apply each merge to pairs of tokens when possible.

    References

    [1] Devlin, Jacob, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. "BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding" Preprint, submitted May 24, 2019. https://doi.org/10.48550/arXiv.1810.04805.

    [2] Wu, Yonghui, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun et al. "Google's Neural Machine Translation System: Bridging the Gap Between Human and Machine Translation." Preprint, submitted October 8, 2016. https://doi.org/10.48550/arXiv.1609.08144

    Version History

    Introduced in R2023b