Develop Reference

Why does tokeniser break down words that are present in vocab

In my understanding, what tokeniser does is that, given each word, the tokeniser will break down the word into sub-words only if the word is not present in the tokeniser.get_vocab() :
def checkModel(model):
tokenizer = AutoTokenizer.from_pretrained(model)
allList = []
for word in tokenizer.get_vocab():
word = word.lower()
tokens = tokenizer.tokenize(word)
try:
if word[0]!='#' and word[0]!='[' and tokens[0] != word:
allList.append((word, tokens))
print(word, tokens)
except:
continue
return allList
checkModel('bert-base-uncased')
# ideally should return an empty list
However, what I have observed is that some models on huggingface will break down words into smaller pieces even if the word is present in the vocab.
checkModel('emilyalsentzer/Bio_ClinicalBERT')
output:
welles ['well', '##es']
lexington ['le', '##xing', '##ton']
palestinian ['pale', '##st', '##inian']
...
elisabeth ['el', '##isa', '##beth']
alexander ['ale', '##xa', '##nder']
appalachian ['app', '##ala', '##chia', '##n']
mitchell ['mit', '##chel', '##l']
...
4630 # tokens in vocab got broken down, not supposed to happen
I have checked a few models of this behaviour, was wondering why is this happening?

This is a really interesting question, and I am currently wondering whether it should be considered as a bug report on the Huggingface repo.
EDIT: I realized that it is possible to define model-specific tokenization_config.json files to overwrite the default behavior. One example is the bert-base-cased repository, which has the following content for the tokenizer config:
{
"do_lower_case": false
}
Given that this functionality is available, I think the best option would be to contact the original author of the work and ask them to potentially consider this configuration (if appropriate for the general use case).
Original Answer:
As it turns out, the vocabulary word that you are checking for is welles, yet the vocab file itself only contains Welles. Notice the difference in the uppercased first letter?
It turns out you can manually force the tokenizer to specifically check for cased vocabulary words, in which case it works fine.
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("emilyalsentzer/Bio_ClinicalBERT",
do_lower_case=False) # This is different
print(tokenizer.do_lower_case)
# Output: False
# Lowercase input will result in split word
tokenizer.convert_ids_to_tokens(tokenizer("welles")["input_ids"])
# Output: ['[CLS]', 'well', '##es', '[SEP]']
# Uppercase input will correctly *not split* the word
tokenizer2.convert_ids_to_tokens(tokenizer2("Welles")["input_ids"])
['[CLS]', 'Welles', '[SEP]']
Per default, however, this is not the case, and all words will be converted to lowercase, which is why you cannot find the word:
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("emilyalsentzer/Bio_ClinicalBERT")
# Per default, lowercasing is enabled!
print(tokenizer.do_lower_case)
# Output: True
# This time now we get the same (lowercased) output both times!
tokenizer.convert_ids_to_tokens(tokenizer("welles")["input_ids"])
['[CLS]', 'well', '##es', '[SEP]']
tokenizer.convert_ids_to_tokens(tokenizer("Welles")["input_ids"])
['[CLS]', 'well', '##es', '[SEP]']

The tokenizer you are calling 'emilyalsentzer/Bio_ClinicalBERT' has tokens that are not present in the original base tokenizer. To add tokens to the tokenizer one can either provide a list of strings or a list of tokenizers.AddedTokens.
The default behavior in both cases is to allow new words to be used as subwords. In my example if we add 'director' and 'cto' to the tokenizer, then 'director' can be broken down into 'dire' + 'cto' + 'r' ('dire' and 'r' are a part of the original tokenizer). To avoid this, one should use:
tokenizer.add_tokens([tokenizers.AddedToken(new_word, single_word = True) for new_word in new_words])
I do think a lot of users would simply use a list of strings (as I did, until half an hour ago). But this would lead to the problem that you saw.
To change this for a customized tokenizer (like 'emilyalsentzer/Bio_ClinicalBERT') w/o losing much in model performance, I'd recommend extracting the set of words from this tokenizer, and comparing it to its base tokenizer (for example 'bert-base-uncased'). This will give you the set of words that were added to the base tokenizer as part of model re-training. Then take the base tokenizer and add this new words to it using AddedToken with single_word set to True. Replace the custom tokenizer with this new tokenizer.

tensorflow_hub to pull BERT embedding on windows machine - extending to albert

Recently I posted this question and tried to solve my problem. My questions are
is my approach correct?
My example sentences length are 7 and 6 respectively - (['New Delhi is the capital of India', 'The capital of India is Delhi']), even if I add cls and sep tokens, the lengths are 9 and 8. max_seq_len parameter is 10, then why the last row of x1 and x2 are not the same?
How to get embedding when I have a paragraph of more than 2 sentences? do i have to pass one sentence at a time? But in such case wont i loose information as I am not passing all sentences together?
I did some additional research and it seems that I can pass entire paragraph as a single sentence using segment_ids as 0 for all words in a paragraph. Is that correct?
how to get embedding for ALBERT? I see that the ALBERT also has tokenization.py file. But I dont see vocab.txt. I see file 30k-clean.vocab. Could i use 30k-clean.vocab instead of vocab.txt?

#user2543622, you may refer to the official code here, in your case, you can do something like:
import tensorflow_hub as hub
albert_module = hub.Module("https://tfhub.dev/google/albert_base/2", trainable=True)
print(albert_module.get_signature_names()) # should output ['tokens', 'tokenization_info', 'mlm']
# then
tokenization_info = albert_module(signature="tokenization_info",
as_dict=True)
with tf.Session() as sess:
vocab_file, do_lower_case = sess.run([tokenization_info["vocab_file"],
tokenization_info["do_lower_case"]])
print(vocab_file) # output b'/var/folders/v6/vnz79w0d2dn95fj0mtnqs27m0000gn/T/tfhub_modules/098d91f064a4f53dffc7633d00c3d8e87f3a4716/assets/30k-clean.model'
I guess this vocab_file is a binary sentencepiece model file, so you should this one for tokenization as below, instead of using the 30k-clean.vocab.
# you still need the tokenization.py code to perform full tokenization
return tokenization.FullTokenizer(
vocab_file=vocab_file, do_lower_case=do_lower_case,
spm_model_file=FLAGS.spm_model_file)
If you only need the embedding matrix values, you take a look at the albert_module.variable_map, e.g.:
print(albert_module.variable_map['bert/embeddings/word_embeddings'])
# <tf.Variable 'module/bert/embeddings/word_embeddings:0' shape=(30000, 128) dtype=float32>

Your approach seems right
Could you please check the tokenizations of sentence 1 and 2 using the tokenizer, this should reveal if there are additional word pieces in one of the sentences. This can be checked as below:
import tokenization
tokenizer = tokenization.FullTokenizer(vocab_file=<PATH to Vocab file>, do_lower_case=True)
tokens = tokenizer.tokenize(example.text_a)
print(tokens)
This should give you word piece tokenized list, without [CLS] and [SEP] token.
Generally, word piece tokenization splits the words when words are not in vocabulary, this would create higher length of tokens than the number of inputs tokens.
You can pass both the sentences together, provided that the length of the paragraph after word piece tokenization does not exceed max_sequence length.
The vocab file for albert is in ./data/vocab.txt directory. Provided you have got the albert code from: here.
In case, if you have got the model from tf-hub, the file is 2/assets/30k-clean.vocab

Why can't spacy differentiate between two homograph tokens in the following code?

A homograph is a word that has the same spelling as another word but has a different sound and a different meaning, for example,lead (to go in front of) / lead (a metal) .
I was trying to use spacy word vectors to compare documents with each other by summing each word vector for each document and then finally finding cosine similarity. If for example spacy vectors have the same vector for the two 'lead' listed above , the results will be probably bad.
In the code below , why does the similarity between the two 'bank'
tokens come out as 1.00 ?
import spacy
nlp = spacy.load('en')
str1 = 'The guy went inside the bank to take out some money'
str2 = 'The house by the river bank.'
str1_tokenized = nlp(str1.decode('utf8'))
str2_tokenized = nlp(str2.decode('utf8'))
token1 = str1_tokenized[-6]
token2 = str2_tokenized[-2]
print 'token1 = {} token2 = {}'.format(token1,token2)
print token1.similarity(token2)
The output for given program is
token1 = bank token2 = bank
1.0

As kntgu already pointed out, spaCy distinguishes tokens by their characters, not by their semantic meaning. The sense2vec approach by the developers of spaCy concatenates tokens with their POS-tag and can help in the case of 'lead_VERB' vs. 'lead_NOUN'. However, it will not help in your example of 'bank (river bank)' vs. 'bank (financial institute)', as both are nouns.
SpaCy does not support any solution to this out of the box, but you can have a look at contextualized word representations like ELMo or BERT. Both generate word vectors for a given sentence, taking the context into account. Therefore, I assume the vectors for both 'bank' tokens will be substantially different.
Both are relatively recent approaches and are not as comfortable to use, but might help in your use case. For ELMo, there is a command line tool which lets you generate word embeddings for a set of sentences without having to write any code: https://github.com/allenai/allennlp/blob/master/tutorials/how_to/elmo.md#writing-contextual-representations-to-disk

Using gensim's Phraser with pre-trained vectors

I am trying to use pre-trained word embeddings taking into account phrases. Popular pre-trained embeddings like GoogleNews-vectors-negative300.bin.gz have separate embeddings for phrases as well as unigrams e.g., embeddings for New_York and the two unigrams New and York. Naive word tokenization and dictionary look-up ignore the bigram embedding.
Gensim provides a nice Phrase model, where given a text sequence it can learn compact phrases e.g., New_York instead of two unigrams New and York. This is done by aggregating and comparing count statistics between the unigrams and the bigram. 1. Is it possible to use Phrase with pre-trained embeddings without estimating the count statistics elsewhere?
Is it possible to use Phrase with pre-trained embeddings without estimating the count statistics elsewhere?
If not, is there an efficient way to use these bigrams? I can imagine a way using a loop, but I believe it is ugly (Below).
Here is the ugly code.
from ntlk import word_tokenize
last_added = False
sentence = 'I love New York.'
tokens = ["<s>"]+ word_tokenize(sentence) +"<\s>"]
vectors = []
for index, token in enumerate(tokens):
if last_added:
last_added=False
continue
if "%s_%s"%(tokens[index-1], token) in model:
vectors.append("%s_%s"%(tokens[index-1], token))
last_added = True
else:
vectors.append(tokens[index-1])
lase_added = False

Probability tree for sentences in nltk employing both lookahead and lookback dependencies

Does nltk or any other NLP tool allow to construct probability trees based on input sentences thus storing the language model of the input text in a dictionary tree, the following example gives the rough idea, but I need the same functionality such that a word Wt does not just probabilistically modelled on past input words(history) Wt-n but also on lookahead words like Wt+m. Also the lookback and lookahead word count should also be 2 or more i.e. bigrams or more. Are there any other libraries in python which achieve this?
from collections import defaultdict
import nltk
import math
ngram = defaultdict(lambda: defaultdict(int))
corpus = "The cat is cute. He jumps and he is happy."
for sentence in nltk.sent_tokenize(corpus):
tokens = map(str.lower, nltk.word_tokenize(sentence))
for token, next_token in zip(tokens, tokens[1:]):
ngram[token][next_token] += 1
for token in ngram:
total = math.log10(sum(ngram[token].values()))
ngram[token] = {nxt: math.log10(v) - total for nxt, v in ngram[token].items()}
the solution requires both lookahead and lookback and a specially sub classed dictionary may help in solving this problem. Can also point to relevant resources which talk about implementing such a system. A nltk.models seemed to be doing something similar but is no longer available. Are there any existing design patterns in NLP which implement this idea? skip gram based models are similar to this idea too but I feel this has should have been implemented already somewhere.

If I understand your question correctly, you are looking for a way to predict the probability of a word given its surrounding context (not just backward context but also the forward context).
One quick hack for your purpose is to train two different language models. One from right to left and the other from left to right and then probability of a word given its context would be the normalized sum of both forward and backward contexts.
Extending your code:
from collections import defaultdict
import nltk
from nltk.tokenize import word_tokenize
import numpy as np
ngram = defaultdict(lambda: defaultdict(int))
ngram_rev = defaultdict(lambda: defaultdict(int)) #reversed n-grams
corpus = "The cat is cute. He jumps and he is happy."
for sentence in nltk.sent_tokenize(corpus):
tokens = map(str.lower, nltk.word_tokenize(sentence))
for token, next_token in zip(tokens, tokens[1:]):
ngram[token][next_token] += 1
for token, rev_token in zip(tokens[1:], tokens):
ngram_rev[token][rev_token] += 1
for token in ngram:
total = np.log(np.sum(ngram[token].values()))
total_rev = np.log(np.sum(ngram_rev[token].values()))
ngram[token] = {nxt: np.log(v) - total
for nxt, v in ngram[token].items()}
ngram_rev[token] = {prv: np.log(v) - total_rev
for prv, v in ngram_rev[token].items()}
Now the context is in both ngram and ngram_rev which respectively hold the forward and backward contexts.
You should also account for smoothing. That is if a given phrase is not seen in your training corpus, you would just get zero probabilities. In order to avoid that, there are many smoothing techniques the most simple of which is the add-on smoothing.

The normal ngram algorithm traditionally works with prior context only, and for good reason: A bigram tagger makes decisions by considering the tags of the last two words, plus the current word. So unless you tag in two passes, the tag of the next word is not yet known. But you are interested in word ngrams, not tag ngrams, so nothing keeps you from training an ngram tagger where the ngram consists of words from both sides. And you can indeed do it easily with the NLTK.
The NLTK's ngram taggers all make tag ngrams, from the left; but you can easily derive your own tagger from their abstract base class, ContextTagger:
import nltk
from nltk.tag import ContextTagger
class TwoSidedTagger(ContextTagger):
left = 2
right = 1
def context(self, tokens, index, history):
left = self.left
right = self.right
tokens = tuple(t.lower() for t in tokens)
if index < left:
tokens = ("<start>",) * left + tokens
index += left
if index + right >= len(tokens):
tokens = tokens + ("<end>",) * right
return tokens[index-left:index+right+1]
This defines a tetragram tagger (2+1+1) where the current word is third in the ngram, not last as usual. You can then initialize and train a tagger just like the regular ngram taggers (see chapter 5 of the NLTK book, especially sections 5.4ff). Let's see first how you'd build a part-of-speech tagger, using a portion of the Brown corpus as training data:
data = list(nltk.corpus.brown.tagged_sents(categories="news"))
train_sents = data[400:]
test_sents = data[:400]
twosidedtagger = TwoSidedTagger({}, backoff=nltk.DefaultTagger('NN'))
twosidedtagger._train(train_sents)
Like all ngram taggers in the NLTK, this one will delegate to the backoff tagger if it is asked to tag an ngram it did not see during training.
For simplicity I used a simple "default tagger" as the backoff tagger, but you'll probably need to use something more powerful (see the NLTK chapter again).
You can then use your tagger to tag new text, or evaluate it with an already tagged test set:
>>> print(twosidedtagger.tag("There were dogs everywhere .".split()))
>>> print(twosidedtagger.evaluate(test_sents))
Predicting words:
The above tagger assigns a POS tag by considering nearby words; but your goal is to predict the word itself, so you need different training data and a different default tagger. The NLTK API expects training data in the form (word, LABEL), where LABEL is the value you want to generate. In your case, LABEL is just the current word itself; so make your training data as follows:
data = [ zip(s,s) for s in nltk.corpus.brown.sents(categories="news") ]
train_sents = data[400:]
test_sents = data[:400]
twosidedtagger = TwoSidedTagger({}, backoff=nltk.DefaultTagger('the')) # most common word
twosidedtagger._train(train_sents)
It makes no sense for the target word to appear in the "context" ngram, so you should also modify the method context() so that the returned ngram does not include it:
def context(self, tokens, index, history):
...
return tokens[index-left:index] + tokens[index+1:index+right+1]
This tagger uses trigrams consisting of two words from the left and one from the right of the current word.
With these modifications, you'll build a tagger that outputs the most likely word at any position. Try it and how you like it.
Prediction:
My expectation is that you'll need a humongous amount of training data before you can get decent performance. The problem is that ngram taggers can only suggest a tag for contexts they saw during training.
To build a tagger that generalizes, consider using the NLTK to train a "sequential classifier". You can use whatever features you want, including the words before and after-- of course, how well it will work is your problem. The NLTK classifier API is similar to that for the ContextTagger, but the context function (aka feature function) returns a dictionary, not a tuple. Again, see the NLTK book and the source code.