Tokens to Words mapping in the tokenizer decode step huggingface? - pytorch

Question

Is there a way to know the mapping from the tokens back to the original words in the tokenizer.decode() function?
For example:
from transformers.tokenization_roberta import RobertaTokenizer
tokenizer = RobertaTokenizer.from_pretrained('roberta-large', do_lower_case=True)
str = "This is a tokenization example"
tokenized = tokenizer.tokenize(str)
## ['this', 'Ġis', 'Ġa', 'Ġtoken', 'ization', 'Ġexample']
encoded = tokenizer.encode_plus(str)
## encoded['input_ids']=[0, 42, 16, 10, 19233, 1938, 1246, 2]
decoded = tokenizer.decode(encoded['input_ids'])
## '<s> this is a tokenization example</s>'
And the objective is to have a function that maps each token in the decode process to the correct input word, for here it will be:
desired_output = [[1],[2],[3],[4,5],[6]] As this corresponds to id 42, while token and ization corresponds to ids [19244,1938] which are at indexes 4,5 of the input_ids array.

Answer 1

As far as I know their is no built-in method for that, but you can create one by yourself:
from transformers.tokenization_roberta import RobertaTokenizer
tokenizer = RobertaTokenizer.from_pretrained('roberta-large', do_lower_case=True)
example = "This is a tokenization example"
print({x : tokenizer.encode(x, add_special_tokens=False, add_prefix_space=True) for x in example.split()})
Output:
{'This': [42], 'is': [16], 'a': [10], 'tokenization': [19233, 1938], 'example': [1246]}
To get exactly your desired output, you have to work with a list comprehension:
#start index because the number of special tokens is fixed for each model (but be aware of single sentence input and pairwise sentence input)
idx = 1
enc =[tokenizer.encode(x, add_special_tokens=False, add_prefix_space=True) for x in example.split()]
desired_output = []
for token in enc:
tokenoutput = []
for ids in token:
tokenoutput.append(idx)
idx +=1
desired_output.append(tokenoutput)
print(desired_output)
Output:
[[1], [2], [3], [4, 5], [6]]

Answer 2

If you use the fast tokenizers, i.e. the rust backed versions from the tokenizers library the encoding contains a word_ids method that can be used to map sub-words back to their original word. What constitutes a word vs a subword depends on the tokenizer, a word is something generated by the pre-tokenization stage, i.e. split by whitespace, a subword is generated by the actual model (BPE or Unigram for example).
The code below should work in general, even if the pre-tokenization performs additional splitting. For example I created my own custom step that splits based on PascalCase - the words here are Pascal and Case, the accepted answer wont work in this case since it assumes words are whitespace delimited.
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('roberta-large', do_lower_case=True)
example = "This is a tokenization example"
encoded = tokenizer(example)
desired_output = []
for word_id in encoded.word_ids():
if word_id is not None:
start, end = encoded.word_to_tokens(word_id)
if start == end - 1:
tokens = [start]
else:
tokens = [start, end-1]
if len(desired_output) == 0 or desired_output[-1] != tokens:
desired_output.append(tokens)
desired_output