R-help-chat / mods /bm25s_tokenization.py
jedick
Don't import tqdm for BM25S tokenizer used in retrieval
142bd00
from ast import Tuple
from pathlib import Path
import re
from typing import Any, Dict, List, Union, Callable, NamedTuple
import typing
from bm25s.utils import json_functions
try:
# To hide progress bars, don't import tqdm
# from tqdm.auto import tqdm
raise ImportError("Not importing tqdm")
except ImportError:
def tqdm(iterable, *args, **kwargs):
return iterable
from bm25s.stopwords import (
STOPWORDS_EN,
STOPWORDS_EN_PLUS,
STOPWORDS_GERMAN,
STOPWORDS_DUTCH,
STOPWORDS_FRENCH,
STOPWORDS_SPANISH,
STOPWORDS_PORTUGUESE,
STOPWORDS_ITALIAN,
STOPWORDS_RUSSIAN,
STOPWORDS_SWEDISH,
STOPWORDS_NORWEGIAN,
STOPWORDS_CHINESE,
)
class Tokenized(NamedTuple):
"""
NamedTuple with two fields: ids and vocab. The ids field is a list of list of token IDs
for each document. The vocab field is a dictionary mapping tokens to their index in the
vocabulary.
"""
ids: List[List[int]]
vocab: Dict[str, int]
def __repr__(self):
"""
Returns:
a string representation of the class.
for example, for a small corpus, it would be something like:
----
Tokenized(
"ids": [
0: [0, 1, 2, 3]
],
"vocab": [
'': 4
'cat': 0
'feline': 1
'likes': 2
'purr': 3
],
)
----
and, for example, for a large corpus, it would be something like:
----
Tokenized(
"ids": [
0: [0, 1, 2, 3]
1: [4, 5, 6, 7, 8, 9]
2: [10, 11, 12, 13, 14]
3: [15, 16, 17, 18, 19]
4: [0, 1, 2, 3, 0, 20, 21, 22, 23, 24, ...]
5: [0, 1, 2, 3]
6: [4, 5, 6, 7, 8, 9]
7: [10, 11, 12, 13, 14]
8: [15, 16, 17, 18, 19]
9: [0, 1, 2, 3, 0, 20, 21, 22, 23, 24, ...]
... (total 500000 docs)
],
"vocab": [
'': 29
'animal': 12
'beautiful': 11
'best': 6
'bird': 10
'can': 13
'carefully': 27
'casually': 28
'cat': 0
'creature': 16
... (total 30 tokens)
],
)
----
"""
lines_print_max_num = 10
single_doc_print_max_len = 10
lines = ["Tokenized(", ' "ids": [']
for doc_idx, document in enumerate(self.ids[:lines_print_max_num]):
preview = document[:single_doc_print_max_len]
if len(document) > single_doc_print_max_len:
preview += ["..."]
lines.append(f" {doc_idx}: [{', '.join([str(x) for x in preview])}]")
if len(self.ids) > lines_print_max_num:
lines.append(f" ... (total {len(self.ids)} docs)")
lines.append(f' ],\n "vocab": [')
vocab_keys = sorted(list(self.vocab.keys()))
for vocab_idx, key_ in enumerate(vocab_keys[:lines_print_max_num]):
val_ = self.vocab[key_]
lines.append(f" {key_!r}: {val_}")
if len(list(vocab_keys)) > 10:
lines.append(f" ... (total {len(vocab_keys)} tokens)")
lines.append(" ],\n)")
return "\n".join(lines)
class Tokenizer:
"""
Tokenizer class for tokenizing a list of strings and converting them to token IDs.
Parameters
----------
lower : bool, optional
Whether to convert the text to lowercase before tokenization
splitter : Union[str, Callable], optional
If a string is provided, the tokenizer will interpret it as a regex pattern,
and use the `re.compile` function to compile the pattern and use the `findall` method
to split the text. If a callable is provided, the tokenizer will use the callable to
split the text. The callable should take a string as input and return a list of strings.
stopwords : Union[str, List[str]], optional
The list of stopwords to remove from the text. If "english" or "en" is provided,
the function will use the default English stopwords. If None or False is provided,
no stopwords will be removed. If a list of strings is provided, the tokenizer will
use the list of strings as stopwords.
stemmer : Callable, optional
The stemmer to use for stemming the tokens. It is recommended
to use the PyStemmer library for stemming, but you can also any callable that
takes a list of strings and returns a list of strings.
"""
def __init__(
self,
lower: bool = True,
splitter: Union[str, Callable] = r"(?u)\b\w\w+\b",
stopwords: Union[str, List[str]] = "english",
stemmer: Callable = None, # type: ignore
):
self.lower = lower
if isinstance(splitter, str):
splitter = re.compile(splitter).findall
if not callable(splitter):
raise ValueError("splitter must be a callable or a regex pattern.")
# Exception handling for stemmer when we are using PyStemmer, which has a stemWords method
if hasattr(stemmer, "stemWord"):
stemmer = stemmer.stemWord
if not callable(stemmer) and stemmer is not None:
raise ValueError("stemmer must be callable or have a `stemWord` method.")
self.stopwords = _infer_stopwords(stopwords)
self.splitter = splitter
self.stemmer = stemmer
self.reset_vocab()
def reset_vocab(self):
"""
Reset the vocabulary dictionaries to empty dictionaries, allowing you to
tokenize a new set of texts without reusing the previous vocabulary.
"""
self.word_to_stem = {} # word -> stemmed word, e.g. "apple" -> "appl"
self.stem_to_sid = {} # stem -> stemmed id, e.g. "appl" -> 0
# word -> {stemmed, unstemmed} id, e.g. "apple" -> 0 (appl) or "apple" -> 2 (apple)
self.word_to_id = {}
def save_vocab(self, save_dir: str, vocab_name: str = "vocab.tokenizer.json"):
"""
Save the vocabulary dictionaries to a file. The file is saved in JSON format.
Parameters
----------
save_dir : str
The directory where the vocabulary file is saved.
vocab_name : str, optional
The name of the vocabulary file. Default is "vocab.tokenizer.json". Make
sure to not use the same name as the vocab.index.json file saved by the BM25
model, as it will overwrite the vocab.index.json file and cause errors.
"""
save_dir: Path = Path(save_dir)
path = save_dir / vocab_name
save_dir.mkdir(parents=True, exist_ok=True)
with open(path, "w", encoding="utf-8") as f:
d = {
"word_to_stem": self.word_to_stem,
"stem_to_sid": self.stem_to_sid,
"word_to_id": self.word_to_id,
}
f.write(json_functions.dumps(d, ensure_ascii=False))
def load_vocab(self, save_dir: str, vocab_name: str = "vocab.tokenizer.json"):
"""
Load the vocabulary dictionaries from a file. The file should be saved in JSON format.
Parameters
----------
save_dir : str
The directory where the vocabulary file is saved.
vocab_name : str, optional
The name of the vocabulary file.
Note
----
The vocabulary file should be saved in JSON format, with the following keys:
- word_to_stem: a dictionary mapping words to their stemmed words
- stem_to_sid: a dictionary mapping stemmed words to their stemmed IDs
- word_to_id: a dictionary mapping words to their word
"""
path = Path(save_dir) / vocab_name
with open(path, "r", encoding="utf-8") as f:
d = json_functions.loads(f.read())
self.word_to_stem = d["word_to_stem"]
self.stem_to_sid = d["stem_to_sid"]
self.word_to_id = d["word_to_id"]
def save_stopwords(
self, save_dir: str, stopwords_name: str = "stopwords.tokenizer.json"
):
"""
Save the stopwords to a file. The file is saved in JSON format.
Parameters
----------
save_dir : str
The directory where the stopwords file is saved.
stopwords_name : str, optional
The name of the stopwords file. Default is "stopwords.tokenizer.json".
"""
save_dir: Path = Path(save_dir)
path = save_dir / stopwords_name
save_dir.mkdir(parents=True, exist_ok=True)
with open(path, "w") as f:
f.write(json_functions.dumps(self.stopwords))
def load_stopwords(
self, save_dir: str, stopwords_name: str = "stopwords.tokenizer.json"
):
"""
Load the stopwords from a file. The file should be saved in JSON format.
Parameters
----------
save_dir : str
The directory where the stopwords file is saved.
stopwords_name : str, optional
The name of the stopwords file.
"""
path = Path(save_dir) / stopwords_name
with open(path, "r") as f:
self.stopwords = json_functions.loads(f.read())
def streaming_tokenize(
self,
texts: List[str],
update_vocab: Union[bool, str] = True,
allow_empty: bool = True,
):
"""
Tokenize a list of strings and return a generator of token IDs.
Parameters
----------
texts : List[str]
A list of strings to tokenize.
update_vocab : bool, optional
Whether to update the vocabulary dictionary with the new tokens. If true,
the different dictionaries making up the vocabulary will be updated with the
new tokens. If False, the function will not update the vocabulary. Unless you have
a stemmer and the stemmed word is in the stem_to_sid dictionary. If "never",
the function will never update the vocabulary, even if the stemmed word is in
the stem_to_sid dictionary. Note that update_vocab="if_empty" is not supported
in this method, only in the `tokenize` method.
allow_empty : bool, optional
Whether to allow the splitter to return an empty string. If False, the splitter
will return an empty list, which may cause issues if the tokenizer is not expecting
an empty list. If True, the splitter will return a list with a single empty string.
"""
stopwords_set = set(self.stopwords) if self.stopwords is not None else None
using_stopwords = stopwords_set is not None
using_stemmer = self.stemmer is not None
if allow_empty is True and update_vocab is True and "" not in self.word_to_id:
idx = max(self.word_to_id.values(), default=-1) + 1
self.word_to_id[""] = idx
if using_stemmer:
if "" not in self.word_to_stem:
self.word_to_stem[""] = ""
if "" not in self.stem_to_sid:
self.stem_to_sid[""] = idx
for text in texts:
if self.lower:
text = text.lower()
splitted_words = list(self.splitter(text))
if allow_empty is True and len(splitted_words) == 0:
splitted_words = [""]
doc_ids = []
for word in splitted_words:
if word in self.word_to_id:
wid = self.word_to_id[word]
doc_ids.append(wid)
continue
if using_stopwords and word in stopwords_set:
continue
# We are always updating the word_to_stem mapping since even new
# words that we have never seen before can be stemmed, with the
# possibility that the stemmed ID is already in the stem_to_sid
if using_stemmer:
if word in self.word_to_stem:
stem = self.word_to_stem[word]
else:
stem = self.stemmer(word)
self.word_to_stem[word] = stem
# if the stem is already in the stem_to_sid, we can just use the ID
# and update the word_to_id dictionary, unless update_vocab is "never"
# in which case we skip this word
if update_vocab != "never" and stem in self.stem_to_sid:
sid = self.stem_to_sid[stem]
self.word_to_id[word] = sid
doc_ids.append(sid)
elif update_vocab is True:
sid = len(self.stem_to_sid)
self.stem_to_sid[stem] = sid
self.word_to_id[word] = sid
doc_ids.append(sid)
else:
# if we are not using a stemmer, we can just update the word_to_id
# directly rather than going through the stem_to_sid dictionary
if update_vocab is True and word not in self.word_to_id:
wid = len(self.word_to_id)
self.word_to_id[word] = wid
doc_ids.append(wid)
if len(doc_ids) == 0 and allow_empty is True and "" in self.word_to_id:
doc_ids = [self.word_to_id[""]]
yield doc_ids
def tokenize(
self,
texts: List[str],
update_vocab: Union[bool, str] = "if_empty",
leave_progress: bool = False,
show_progress: bool = True,
length: Union[int, None] = None,
return_as: str = "ids",
allow_empty: bool = True,
) -> Union[List[List[int]], List[List[str]], typing.Generator, Tokenized]:
"""
Tokenize a list of strings and return the token IDs.
Parameters
----------
texts : List[str]
A list of strings to tokenize.
update_vocab : bool, optional
Whether to update the vocabulary dictionary with the new tokens. If true,
the different dictionaries making up the vocabulary will be updated with the
new tokens. If False, the vocabulary will not be updated unless you have a stemmer
and the stemmed word is in the stem_to_sid dictionary. If update_vocab="if_empty",
the function will only update the vocabulary if it is empty, i.e. when the
function is called for the first time, or if the vocabulary has been reset with
the `reset_vocab` method. If update_vocab="never", the "word_to_id" will never
be updated, even if the stemmed word is in the stem_to_sid dictionary. Only use
this if you are sure that the stemmed words are already in the stem_to_sid dictionary.
leave_progress : bool, optional
Whether to leave the progress bar after completion. If False, the progress bar
will disappear after completion. If True, the progress bar will stay on the screen.
show_progress : bool, optional
Whether to show the progress bar for tokenization. If False, the function will
not show the progress bar. If True, it will use tqdm.auto to show the progress bar.
length : int, optional
The length of the texts. If None, the function will call `len(texts)` to get the length.
This is mainly used when `texts` is a generator or a stream instead of a list, in which case
`len(texts)` will raise a TypeError, and you need to provide the length manually.
return_as : str, optional
The type of object to return by this function.
If "tuple", this returns a Tokenized namedtuple, which contains the token IDs
and the vocab dictionary.
If "string", this return a list of lists of strings, each string being a token.
If "ids", this return a list of lists of integers corresponding to the token IDs,
or stemmed IDs if a stemmer is used.
allow_empty : bool, optional
Whether to allow the splitter to return an empty string. If False, the splitter
will return an empty list, which may cause issues if the tokenizer is not expecting
an empty list. If True, the splitter will return a list with a single empty string.
Returns
-------
List[List[int]] or Generator[List[int]] or List[List[str]] or Tokenized object
If `return_as="stream"`, a Generator[List[int]] is returned, each integer being a token ID.
If `return_as="ids"`, a List[List[int]] is returned, each integer being a token ID.
If `return_as="string"`, a List[List[str]] is returned, each string being a token.
If `return_as="tuple"`, a Tokenized namedtuple is returned, with names `ids` and `vocab`.
"""
incorrect_return_error = (
"return_as must be either 'tuple', 'string', 'ids', or 'stream'."
)
incorrect_update_vocab_error = (
"update_vocab must be either True, False, 'if_empty', or 'never'."
)
if return_as not in ["tuple", "string", "ids", "stream"]:
raise ValueError(incorrect_return_error)
if update_vocab not in [True, False, "if_empty", "never"]:
raise ValueError(incorrect_update_vocab_error)
if update_vocab == "if_empty":
update_vocab = len(self.word_to_id) == 0
stream_fn = self.streaming_tokenize(
texts=texts, update_vocab=update_vocab, allow_empty=allow_empty
)
if return_as == "stream":
return stream_fn
if length is None:
length = len(texts)
tqdm_kwargs = dict(
desc="Tokenize texts",
leave=leave_progress,
disable=not show_progress,
total=length,
)
token_ids = []
for doc_ids in tqdm(stream_fn, **tqdm_kwargs):
token_ids.append(doc_ids)
if return_as == "ids":
return token_ids
elif return_as == "string":
return self.decode(token_ids)
elif return_as == "tuple":
return self.to_tokenized_tuple(token_ids)
else:
raise ValueError(incorrect_return_error)
def get_vocab_dict(self) -> Dict[str, Any]:
if self.stemmer is None:
# if we are not using a stemmer, we return the word_to_id dictionary
# which maps the words to the word IDs
return self.word_to_id
else:
# if we are using a stemmer, we return the stem_to_sid dictionary,
# which we will use to map the stemmed words to the stemmed IDs
return self.stem_to_sid
def to_tokenized_tuple(self, docs: List[List[int]]) -> Tokenized:
"""
Convert the token IDs to a Tokenized namedtuple, which contains the word IDs, or the stemmed IDs
if a stemmer is used. The Tokenized namedtuple contains two fields: ids and vocab. The latter
is a dictionary mapping the token IDs to the tokens, or a dictionary mapping the stemmed IDs to
the stemmed tokens (if a stemmer is used).
"""
return Tokenized(ids=docs, vocab=self.get_vocab_dict())
def decode(self, docs: List[List[int]]) -> List[List[str]]:
"""
Convert word IDs (or stemmed IDs if a stemmer is used) back to strings using the vocab dictionary,
which is a dictionary mapping the word IDs to the words or a dictionary mapping the stemmed IDs
to the stemmed words (if a stemmer is used).
Parameters
----------
docs : List[List[int]]
A list of lists of word IDs or stemmed IDs.
Returns
-------
List[List[str]]
A list of lists of strings, each string being a word or a stemmed word if a stemmer is used.
"""
vocab = self.get_vocab_dict()
reverse_vocab = {v: k for k, v in vocab.items()}
return [[reverse_vocab[token_id] for token_id in doc] for doc in docs]
def convert_tokenized_to_string_list(tokenized: Tokenized) -> List[List[str]]:
"""
Convert the token IDs back to strings using the vocab dictionary.
"""
reverse_vocab = {v: k for k, v in tokenized.vocab.items()}
return [
[reverse_vocab[token_id] for token_id in doc_ids] for doc_ids in tokenized.ids
]
def _infer_stopwords(stopwords: Union[str, List[str]]) -> Union[List[str], tuple]:
# Source of stopwords: https://github.com/nltk/nltk/blob/96ee715997e1c8d9148b6d8e1b32f412f31c7ff7/nltk/corpus/__init__.py#L315
if stopwords in ["english", "en", True]: # True is added to support the default
return STOPWORDS_EN
elif stopwords in ["english_plus", "en_plus"]:
return STOPWORDS_EN_PLUS
elif stopwords in ["german", "de"]:
return STOPWORDS_GERMAN
elif stopwords in ["dutch", "nl"]:
return STOPWORDS_DUTCH
elif stopwords in ["french", "fr"]:
return STOPWORDS_FRENCH
elif stopwords in ["spanish", "es"]:
return STOPWORDS_SPANISH
elif stopwords in ["portuguese", "pt"]:
return STOPWORDS_PORTUGUESE
elif stopwords in ["italian", "it"]:
return STOPWORDS_ITALIAN
elif stopwords in ["russian", "ru"]:
return STOPWORDS_RUSSIAN
elif stopwords in ["swedish", "sv"]:
return STOPWORDS_SWEDISH
elif stopwords in ["norwegian", "no"]:
return STOPWORDS_NORWEGIAN
elif stopwords in ["chinese", "zh"]:
return STOPWORDS_CHINESE
elif stopwords in [None, False]:
return []
elif isinstance(stopwords, str):
raise ValueError(
f"{stopwords} not recognized. Only English stopwords as default, German, Dutch, French, Spanish, Portuguese, Italian, Russian, Swedish, Norwegian, and Chinese are currently supported. "
"Please input a list of stopwords"
)
else:
return stopwords
def tokenize(
texts: Union[str, List[str]],
lower: bool = True,
token_pattern: str = r"(?u)\b\w\w+\b",
stopwords: Union[str, List[str]] = "english",
stemmer: Callable = None, # type: ignore
return_ids: bool = True,
show_progress: bool = True,
leave: bool = False,
allow_empty: bool = True,
) -> Union[List[List[str]], Tokenized]:
"""
Tokenize a list using the same method as the scikit-learn CountVectorizer,
and optionally apply a stemmer to the tokens or stopwords removal.
If you provide stemmer, it must have a `stemWords` method, or be callable
that takes a list of strings and returns a list of strings. If your stemmer
can only be called on a single word, you can use a lambda function to wrap it,
e.g. `lambda lst: list(map(stemmer.stem, lst))`.
If return_ids is True, the function will return a namedtuple with: (1) the tokenized
IDs and (2) the token_to_index dictionary. You can access the tokenized IDs using
the `ids` attribute and the token_to_index dictionary using the `vocab` attribute,
You can also destructure the namedtuple to get the ids and vocab_dict variables,
e.g. `token_ids, vocab = tokenize(...)`.
Parameters
----------
texts : Union[str, List[str]]
A list of strings to tokenize. If a single string is provided, it will be
converted to a list with a single element.
lower : bool, optional
Whether to convert the text to lowercase before tokenization
token_pattern : str, optional
The regex pattern to use for tokenization, by default, r"(?u)\\b\\w\\w+\\b"
stopwords : Union[str, List[str]], optional
The list of stopwords to remove from the text. If "english" or "en" is provided,
the function will use the default English stopwords
stemmer : Callable, optional
The stemmer to use for stemming the tokens. It is recommended
to use the PyStemmer library for stemming, but you can also any callable that
takes a list of strings and returns a list of strings.
return_ids : bool, optional
Whether to return the tokenized IDs and the vocab dictionary. If False, the
function will return the tokenized strings. If True, the function will return
a namedtuple with the tokenized IDs and the vocab dictionary.
show_progress : bool, optional
Whether to show the progress bar for tokenization. If False, the function will
not show the progress bar. If True, it will use tqdm.auto to show the progress bar.
leave : bool, optional
Whether to leave the progress bar after completion. If False, the progress bar
will disappear after completion. If True, the progress bar will stay on the screen.
allow_empty : bool, optional
Whether to allow the splitter to return an empty string. If False, the splitter
will return an empty list, which may cause issues if the tokenizer is not expecting
an empty list. If True, the splitter will return a list with a single empty string.
Note
-----
You may pass a single string or a list of strings. If you pass a single string,
this function will convert it to a list of strings with a single element.
"""
if isinstance(texts, str):
texts = [texts]
split_fn = re.compile(token_pattern).findall
stopwords = _infer_stopwords(stopwords)
# Step 1: Split the strings using the regex pattern
corpus_ids = []
token_to_index = {}
for text in tqdm(
texts, desc="Split strings", leave=leave, disable=not show_progress
):
stopwords_set = set(stopwords)
if lower:
text = text.lower()
splitted = split_fn(text)
if allow_empty is False and len(splitted) == 0:
splitted = [""]
doc_ids = []
for token in splitted:
if token in stopwords_set:
continue
if token not in token_to_index:
token_to_index[token] = len(token_to_index)
token_id = token_to_index[token]
doc_ids.append(token_id)
corpus_ids.append(doc_ids)
# Create a list of unique tokens that we will use to create the vocabulary
unique_tokens = list(token_to_index.keys())
# Step 2: Stem the tokens if a stemmer is provided
if stemmer is not None:
if hasattr(stemmer, "stemWords"):
stemmer_fn = stemmer.stemWords
elif callable(stemmer):
stemmer_fn = stemmer
else:
error_msg = "Stemmer must have a `stemWord` method, or be callable. For example, you can use the PyStemmer library."
raise ValueError(error_msg)
# Now, we use the stemmer on the token_to_index dictionary to get the stemmed tokens
tokens_stemmed = stemmer_fn(unique_tokens)
vocab = set(tokens_stemmed)
vocab_dict = {token: i for i, token in enumerate(vocab)}
stem_id_to_stem = {v: k for k, v in vocab_dict.items()}
# We create a dictionary mapping the stemmed tokens to their index
doc_id_to_stem_id = {
token_to_index[token]: vocab_dict[stem]
for token, stem in zip(unique_tokens, tokens_stemmed)
}
# Now, we simply need to replace the tokens in the corpus with the stemmed tokens
for i, doc_ids in enumerate(
tqdm(corpus_ids, desc="Stem Tokens", leave=leave, disable=not show_progress)
):
corpus_ids[i] = [doc_id_to_stem_id[doc_id] for doc_id in doc_ids]
else:
vocab_dict = token_to_index
# Step 3: Return the tokenized IDs and the vocab dictionary or the tokenized strings
if return_ids:
return Tokenized(ids=corpus_ids, vocab=vocab_dict)
else:
# We need a reverse dictionary to convert the token IDs back to tokens
reverse_dict = stem_id_to_stem if stemmer is not None else unique_tokens
# We convert the token IDs back to tokens in-place
for i, token_ids in enumerate(
tqdm(
corpus_ids,
desc="Reconstructing token strings",
leave=leave,
disable=not show_progress,
)
):
corpus_ids[i] = [reverse_dict[token_id] for token_id in token_ids]
return corpus_ids