Dear Community Members,
During the pre-processing of data, after splitting the raw_data into tokens, I have used the popular WordNet Lemmatizer to generate the stems. I am performing experiments on a dataset that has 18953 tokens.
My question is, does the lemmatization process reduce the size of corpus? I am confused, kindly help in this regard. Any help is appreciated!
Lemmatization converts each token (aka form) in the sentence into its lemma form (aka type):
>>> from nltk import word_tokenize
>>> from pywsd.utils import lemmatize_sentence
>>> text = ['This is a corpus with multiple sentences.', 'This was the second sentence running.', 'For some reasons, there is a need to second foo bar ran.']
>>> lemmatize_sentence(text[0]) # Lemmatized sentence example.
['this', 'be', 'a', 'corpus', 'with', 'multiple', 'sentence', '.']
>>> word_tokenize(text[0]) # Tokenized sentence example.
['This', 'is', 'a', 'corpus', 'with', 'multiple', 'sentences', '.']
>>> word_tokenize(text[0].lower()) # Lowercased and tokenized sentence example.
['this', 'is', 'a', 'corpus', 'with', 'multiple', 'sentences', '.']
If we lemmatize the sentence, each token should receive the corresponding lemma form, so the no. of "words" remains the same whether it's the form or the type:
>>> num_tokens = sum([len(word_tokenize(sent.lower())) for sent in text])
>>> num_lemmas = sum([len(lemmatize_sentence(sent)) for sent in text])
>>> num_tokens, num_lemmas
(29, 29)
>>> [lemmatize_sentence(sent) for sent in text] # lemmatized sentences
[['this', 'be', 'a', 'corpus', 'with', 'multiple', 'sentence', '.'], ['this', 'be', 'the', 'second', 'sentence', 'running', '.'], ['for', 'some', 'reason', ',', 'there', 'be', 'a', 'need', 'to', 'second', 'foo', 'bar', 'ran', '.']]
>>> [word_tokenize(sent.lower()) for sent in text] # tokenized sentences
[['this', 'is', 'a', 'corpus', 'with', 'multiple', 'sentences', '.'], ['this', 'was', 'the', 'second', 'sentence', 'running', '.'], ['for', 'some', 'reasons', ',', 'there', 'is', 'a', 'need', 'to', 'second', 'foo', 'bar', 'ran', '.']]
The "compression" per-se would refer to the number of unique tokens represented in the whole corpus after you've lemmatized the sentences, e.g.
>>> lemma_vocab = set(chain(*[lemmatize_sentence(sent) for sent in text]))
>>> token_vocab = set(chain(*[word_tokenize(sent.lower()) for sent in text]))
>>> len(lemma_vocab), len(token_vocab)
(21, 23)
>>> lemma_vocab
{'the', 'this', 'to', 'reason', 'for', 'second', 'a', 'running', 'some', 'sentence', 'be', 'foo', 'ran', 'with', '.', 'need', 'multiple', 'bar', 'corpus', 'there', ','}
>>> token_vocab
{'the', 'this', 'to', 'for', 'sentences', 'a', 'second', 'running', 'some', 'is', 'sentence', 'foo', 'reasons', 'with', 'ran', '.', 'need', 'multiple', 'bar', 'corpus', 'there', 'was', ','}
Note: Lemmatization is a pre-processing step. But it should not overwrite your original corpus with the lemmatize forms.