keyphrase-vectorizers
Advanced tools
Set of vectorizers that extract keyphrases with part-of-speech patterns from a collection of text documents and convert them into a document-keyphrase matrix.
This package was developed during the writing of our PatternRank paper. You can check out the paper here. When using KeyphraseVectorizers or PatternRank in academic papers and theses, please use the BibTeX entry below.
Set of vectorizers that extract keyphrases with part-of-speech patterns from a collection of text documents and convert them into a document-keyphrase matrix. A document-keyphrase matrix is a mathematical matrix that describes the frequency of keyphrases that occur in a collection of documents. The matrix rows indicate the text documents and columns indicate the unique keyphrases.
The package contains wrappers of the sklearn.feature_extraction.text.CountVectorizer and sklearn.feature_extraction.text.TfidfVectorizer classes. Instead of using n-gram tokens of a pre-defined range, these classes extract keyphrases from text documents using part-of-speech tags to compute document-keyphrase matrices.
Corresponding medium posts can be found here and here.
First, the document texts are annotated with spaCy part-of-speech tags. A list of
all possible spaCy part-of-speech tags for different languages is
linked here. The annotation
requires passing the spaCy pipeline of the corresponding language
to the vectorizer with the spacy_pipeline parameter.
Second, words are extracted from the document texts whose part-of-speech tags match the regex pattern defined in
the pos_pattern
parameter. The keyphrases are a list of unique words extracted from text documents by this method.
Finally, the vectorizers calculate document-keyphrase matrices.
pip install keyphrase-vectorizers
For detailed information visit the API Guide.
from keyphrase_vectorizers import KeyphraseCountVectorizer
docs = ["""Supervised learning is the machine learning task of learning a function that
maps an input to an output based on example input-output pairs. It infers a
function from labeled training data consisting of a set of training examples.
In supervised learning, each example is a pair consisting of an input object
(typically a vector) and a desired output value (also called the supervisory signal).
A supervised learning algorithm analyzes the training data and produces an inferred function,
which can be used for mapping new examples. An optimal scenario will allow for the
algorithm to correctly determine the class labels for unseen instances. This requires
the learning algorithm to generalize from the training data to unseen situations in a
'reasonable' way (see inductive bias).""",
"""Keywords are defined as phrases that capture the main topics discussed in a document.
As they offer a brief yet precise summary of document content, they can be utilized for various applications.
In an information retrieval environment, they serve as an indication of document relevance for users, as the list
of keywords can quickly help to determine whether a given document is relevant to their interest.
As keywords reflect a document's main topics, they can be utilized to classify documents into groups
by measuring the overlap between the keywords assigned to them. Keywords are also used proactively
in information retrieval."""]
# Init default vectorizer.
vectorizer = KeyphraseCountVectorizer()
# Print parameters
print(vectorizer.get_params())
>>> {'binary': False, 'dtype': <class 'numpy.int64'>, 'lowercase': True, 'max_df': None, 'min_df': None, 'pos_pattern': '<J.*>*<N.*>+', 'spacy_exclude': ['parser', 'attribute_ruler', 'lemmatizer', 'ner'], 'spacy_pipeline': 'en_core_web_sm', 'stop_words': 'english', 'workers': 1}
By default, the vectorizer is initialized for the English language. That means, an English spacy_pipeline is
specified, English stop_words are removed, and the pos_pattern extracts keywords that have 0 or more adjectives,
followed by 1 or more nouns using the English spaCy part-of-speech tags. In addition, the spaCy pipeline
components ['parser', 'attribute_ruler', 'lemmatizer', 'ner'] are excluded by default to increase efficiency. If you
choose a different spacy_pipeline, you may have to exclude/include different pipeline components using
the spacy_exclude parameter for the spaCy POS
tagger to work properly.
# After initializing the vectorizer, it can be fitted
# to learn the keyphrases from the text documents.
vectorizer.fit(docs)
# After learning the keyphrases, they can be returned.
keyphrases = vectorizer.get_feature_names_out()
print(keyphrases)
>>> ['users' 'main topics' 'learning algorithm' 'overlap' 'documents' 'output'
'keywords' 'precise summary' 'new examples' 'training data' 'input'
'document content' 'training examples' 'unseen instances'
'optimal scenario' 'document' 'task' 'supervised learning algorithm'
'example' 'interest' 'function' 'example input' 'various applications'
'unseen situations' 'phrases' 'indication' 'inductive bias'
'supervisory signal' 'document relevance' 'information retrieval' 'set'
'input object' 'groups' 'output value' 'list' 'learning' 'output pairs'
'pair' 'class labels' 'supervised learning' 'machine'
'information retrieval environment' 'algorithm' 'vector' 'way']
# After fitting, the vectorizer can transform the documents
# to a document-keyphrase matrix.
# Matrix rows indicate the documents and columns indicate the unique keyphrases.
# Each cell represents the count.
document_keyphrase_matrix = vectorizer.transform(docs).toarray()
print(document_keyphrase_matrix)
>>> [[0 0 2 0 0 3 0 0 1 3 3 0 1 1 1 0 1 1 2 0 3 1 0 1 0 0 1 1 0 0 1 1 0 1 0 6
1 1 1 3 1 0 3 1 1]
[1 2 0 1 1 0 5 1 0 0 0 1 0 0 0 5 0 0 0 1 0 0 1 0 1 1 0 0 1 2 0 0 1 0 1 0
0 0 0 0 0 1 0 0 0]]
# Fit and transform can also be executed in one step,
# which is more efficient.
document_keyphrase_matrix = vectorizer.fit_transform(docs).toarray()
print(document_keyphrase_matrix)
>>> [[0 0 2 0 0 3 0 0 1 3 3 0 1 1 1 0 1 1 2 0 3 1 0 1 0 0 1 1 0 0 1 1 0 1 0 6
1 1 1 3 1 0 3 1 1]
[1 2 0 1 1 0 5 1 0 0 0 1 0 0 0 5 0 0 0 1 0 0 1 0 1 1 0 0 1 2 0 0 1 0 1 0
0 0 0 0 0 1 0 0 0]]
german_docs = ["""Goethe stammte aus einer angesehenen bürgerlichen Familie.
Sein Großvater mütterlicherseits war als Stadtschultheiß höchster Justizbeamter der Stadt Frankfurt,
sein Vater Doktor der Rechte und Kaiserlicher Rat. Er und seine Schwester Cornelia erfuhren eine aufwendige
Ausbildung durch Hauslehrer. Dem Wunsch seines Vaters folgend, studierte Goethe in Leipzig und Straßburg
Rechtswissenschaft und war danach als Advokat in Wetzlar und Frankfurt tätig.
Gleichzeitig folgte er seiner Neigung zur Dichtkunst.""",
"""Friedrich Schiller wurde als zweites Kind des Offiziers, Wundarztes und Leiters der Hofgärtnerei in
Marbach am Neckar Johann Kaspar Schiller und dessen Ehefrau Elisabetha Dorothea Schiller, geb. Kodweiß,
die Tochter eines Wirtes und Bäckers war, 1759 in Marbach am Neckar geboren
"""]
# Init vectorizer for the german language
vectorizer = KeyphraseCountVectorizer(spacy_pipeline='de_core_news_sm', pos_pattern='<ADJ.*>*<N.*>+', stop_words='german')
The German spacy_pipeline is specified and German stop_words are removed. Because the German spaCy part-of-speech
tags differ from the English ones, the pos_pattern parameter is also customized. The regex pattern <ADJ.*>*<N.*>+
extracts keywords that have 0 or more adjectives, followed by 1 or more nouns using the German spaCy part-of-speech
tags.
Attention! The spaCy pipeline components ['parser', 'attribute_ruler', 'lemmatizer', 'ner'] are excluded by
default to increase efficiency. If you choose a different spacy_pipeline, you may have to exclude/include different
pipeline components using the spacy_exclude parameter for the spaCy POS tagger to work properly.
The KeyphraseTfidfVectorizer has the same function calls and features as the KeyphraseCountVectorizer. The only
difference is, that document-keyphrase matrix cells represent tf or tf-idf values, depending on the parameter settings,
instead of counts.
from keyphrase_vectorizers import KeyphraseTfidfVectorizer
docs = ["""Supervised learning is the machine learning task of learning a function that
maps an input to an output based on example input-output pairs. It infers a
function from labeled training data consisting of a set of training examples.
In supervised learning, each example is a pair consisting of an input object
(typically a vector) and a desired output value (also called the supervisory signal).
A supervised learning algorithm analyzes the training data and produces an inferred function,
which can be used for mapping new examples. An optimal scenario will allow for the
algorithm to correctly determine the class labels for unseen instances. This requires
the learning algorithm to generalize from the training data to unseen situations in a
'reasonable' way (see inductive bias).""",
"""Keywords are defined as phrases that capture the main topics discussed in a document.
As they offer a brief yet precise summary of document content, they can be utilized for various applications.
In an information retrieval environment, they serve as an indication of document relevance for users, as the list
of keywords can quickly help to determine whether a given document is relevant to their interest.
As keywords reflect a document's main topics, they can be utilized to classify documents into groups
by measuring the overlap between the keywords assigned to them. Keywords are also used proactively
in information retrieval."""]
# Init default vectorizer for the English language that computes tf-idf values
vectorizer = KeyphraseTfidfVectorizer()
# Print parameters
print(vectorizer.get_params())
>>> {'binary': False, 'custom_pos_tagger': None, 'decay': None, 'delete_min_df': None, 'dtype': <
class 'numpy.int64'>, 'lowercase': True, 'max_df': None
, 'min_df': None, 'pos_pattern': '<J.*>*<N.*>+', 'spacy_exclude': ['parser', 'attribute_ruler', 'lemmatizer', 'ner',
'textcat'], 'spacy_pipeline': 'en_core_web_sm', 'stop_words': 'english', 'workers': 1}
To calculate tf values instead, set use_idf=False.
# Fit and transform to document-keyphrase matrix.
document_keyphrase_matrix = vectorizer.fit_transform(docs).toarray()
print(document_keyphrase_matrix)
>>> [[0. 0. 0.09245003 0.09245003 0.09245003 0.09245003
0.2773501 0.09245003 0.2773501 0.2773501 0.09245003 0.
0. 0.09245003 0. 0.2773501 0.09245003 0.09245003
0. 0.09245003 0.09245003 0.09245003 0.09245003 0.09245003
0.5547002 0. 0. 0.09245003 0.09245003 0.
0.2773501 0.18490007 0.09245003 0. 0.2773501 0.
0. 0.09245003 0. 0.09245003 0. 0.
0. 0.18490007 0. ]
[0.11867817 0.11867817 0. 0. 0. 0.
0. 0. 0. 0. 0. 0.11867817
0.11867817 0. 0.11867817 0. 0. 0.
0.11867817 0. 0. 0. 0. 0.
0. 0.11867817 0.23735633 0. 0. 0.11867817
0. 0. 0. 0.23735633 0. 0.11867817
0.11867817 0. 0.59339083 0. 0.11867817 0.11867817
0.11867817 0. 0.59339083]]
# Return keyphrases
keyphrases = vectorizer.get_feature_names_out()
print(keyphrases)
>>> ['various applications' 'list' 'task' 'supervisory signal'
'inductive bias' 'supervised learning algorithm' 'supervised learning'
'example input' 'input' 'algorithm' 'set' 'precise summary' 'documents'
'input object' 'interest' 'function' 'class labels' 'machine'
'document content' 'output pairs' 'new examples' 'unseen situations'
'vector' 'output value' 'learning' 'document relevance' 'main topics'
'pair' 'training examples' 'information retrieval environment'
'training data' 'example' 'optimal scenario' 'information retrieval'
'output' 'groups' 'indication' 'unseen instances' 'keywords' 'way'
'phrases' 'overlap' 'users' 'learning algorithm' 'document']
KeyphraseVectorizers loads a spacy.Language object for every KeyphraseVectorizer object.
When using multiple KeyphraseVectorizer objects, it is more efficient to load the spacy.Language object beforehand and pass it as the spacy_pipeline argument.
import spacy
from keyphrase_vectorizers import KeyphraseCountVectorizer, KeyphraseTfidfVectorizer
docs = ["""Supervised learning is the machine learning task of learning a function that
maps an input to an output based on example input-output pairs. It infers a
function from labeled training data consisting of a set of training examples.
In supervised learning, each example is a pair consisting of an input object
(typically a vector) and a desired output value (also called the supervisory signal).
A supervised learning algorithm analyzes the training data and produces an inferred function,
which can be used for mapping new examples. An optimal scenario will allow for the
algorithm to correctly determine the class labels for unseen instances. This requires
the learning algorithm to generalize from the training data to unseen situations in a
'reasonable' way (see inductive bias).""",
"""Keywords are defined as phrases that capture the main topics discussed in a document.
As they offer a brief yet precise summary of document content, they can be utilized for various applications.
In an information retrieval environment, they serve as an indication of document relevance for users, as the list
of keywords can quickly help to determine whether a given document is relevant to their interest.
As keywords reflect a document's main topics, they can be utilized to classify documents into groups
by measuring the overlap between the keywords assigned to them. Keywords are also used proactively
in information retrieval."""]
nlp = spacy.load("en_core_web_sm")
vectorizer1 = KeyphraseCountVectorizer(spacy_pipeline=nlp)
vectorizer2 = KeyphraseTfidfVectorizer(spacy_pipeline=nlp)
# the following calls use the nlp object
vectorizer1.fit(docs)
vectorizer2.fit(docs)
To use a different part-of-speech tagger than the ones provided by spaCy, a custom POS-tagger function can be defined and passed to the KeyphraseVectorizers via the custom_pos_tagger parameter. This parameter expects a callable function which in turn needs to expect a list of strings in a 'raw_documents' parameter and has to return a list of (word token, POS-tag) tuples. If this parameter is not None, the custom tagger function is used to tag words with parts-of-speech, while the spaCy pipeline is ignored.
Flair can be installed via pip install flair.
from typing import List
import flair
from flair.models import SequenceTagger
from flair.tokenization import SegtokSentenceSplitter
docs = ["""Supervised learning is the machine learning task of learning a function that
maps an input to an output based on example input-output pairs. It infers a
function from labeled training data consisting of a set of training examples.
In supervised learning, each example is a pair consisting of an input object
(typically a vector) and a desired output value (also called the supervisory signal).
A supervised learning algorithm analyzes the training data and produces an inferred function,
which can be used for mapping new examples. An optimal scenario will allow for the
algorithm to correctly determine the class labels for unseen instances. This requires
the learning algorithm to generalize from the training data to unseen situations in a
'reasonable' way (see inductive bias).""",
"""Keywords are defined as phrases that capture the main topics discussed in a document.
As they offer a brief yet precise summary of document content, they can be utilized for various applications.
In an information retrieval environment, they serve as an indication of document relevance for users, as the list
of keywords can quickly help to determine whether a given document is relevant to their interest.
As keywords reflect a document's main topics, they can be utilized to classify documents into groups
by measuring the overlap between the keywords assigned to them. Keywords are also used proactively
in information retrieval."""]
# define flair POS-tagger and splitter
tagger = SequenceTagger.load('pos')
splitter = SegtokSentenceSplitter()
# define custom POS-tagger function using flair
def custom_pos_tagger(raw_documents: List[str], tagger: flair.models.SequenceTagger = tagger, splitter: flair.tokenization.SegtokSentenceSplitter = splitter)->List[tuple]:
"""
Important:
The mandatory 'raw_documents' parameter can NOT be named differently and has to expect a list of strings.
Any other parameter of the custom POS-tagger function can be arbitrarily defined, depending on the respective use case.
Furthermore the function has to return a list of (word token, POS-tag) tuples.
"""
# split texts into sentences
sentences = []
for doc in raw_documents:
sentences.extend(splitter.split(doc))
# predict POS tags
tagger.predict(sentences)
# iterate through sentences to get word tokens and predicted POS-tags
pos_tags = []
words = []
for sentence in sentences:
pos_tags.extend([label.value for label in sentence.get_labels('pos')])
words.extend([word.text for word in sentence])
return list(zip(words, pos_tags))
# check that the custom POS-tagger function returns a list of (word token, POS-tag) tuples
print(custom_pos_tagger(raw_documents=docs))
>>> [('Supervised', 'VBN'), ('learning', 'NN'), ('is', 'VBZ'), ('the', 'DT'), ('machine', 'NN'), ('learning', 'VBG'), ('task', 'NN'), ('of', 'IN'), ('learning', 'VBG'), ('a', 'DT'), ('function', 'NN'), ('that', 'WDT'), ('maps', 'VBZ'), ('an', 'DT'), ('input', 'NN'), ('to', 'IN'), ('an', 'DT'), ('output', 'NN'), ('based', 'VBN'), ('on', 'IN'), ('example', 'NN'), ('input-output', 'NN'), ('pairs', 'NNS'), ('.', '.'), ('It', 'PRP'), ('infers', 'VBZ'), ('a', 'DT'), ('function', 'NN'), ('from', 'IN'), ('labeled', 'VBN'), ('training', 'NN'), ('data', 'NNS'), ('consisting', 'VBG'), ('of', 'IN'), ('a', 'DT'), ('set', 'NN'), ('of', 'IN'), ('training', 'NN'), ('examples', 'NNS'), ('.', '.'), ('In', 'IN'), ('supervised', 'JJ'), ('learning', 'NN'), (',', ','), ('each', 'DT'), ('example', 'NN'), ('is', 'VBZ'), ('a', 'DT'), ('pair', 'NN'), ('consisting', 'VBG'), ('of', 'IN'), ('an', 'DT'), ('input', 'NN'), ('object', 'NN'), ('(', ':'), ('typically', 'RB'), ('a', 'DT'), ('vector', 'NN'), (')', ','), ('and', 'CC'), ('a', 'DT'), ('desired', 'VBN'), ('output', 'NN'), ('value', 'NN'), ('(', ','), ('also', 'RB'), ('called', 'VBN'), ('the', 'DT'), ('supervisory', 'JJ'), ('signal', 'NN'), (')', '-RRB-'), ('.', '.'), ('A', 'DT'), ('supervised', 'JJ'), ('learning', 'NN'), ('algorithm', 'NN'), ('analyzes', 'VBZ'), ('the', 'DT'), ('training', 'NN'), ('data', 'NNS'), ('and', 'CC'), ('produces', 'VBZ'), ('an', 'DT'), ('inferred', 'JJ'), ('function', 'NN'), (',', ','), ('which', 'WDT'), ('can', 'MD'), ('be', 'VB'), ('used', 'VBN'), ('for', 'IN'), ('mapping', 'VBG'), ('new', 'JJ'), ('examples', 'NNS'), ('.', '.'), ('An', 'DT'), ('optimal', 'JJ'), ('scenario', 'NN'), ('will', 'MD'), ('allow', 'VB'), ('for', 'IN'), ('the', 'DT'), ('algorithm', 'NN'), ('to', 'TO'), ('correctly', 'RB'), ('determine', 'VB'), ('the', 'DT'), ('class', 'NN'), ('labels', 'NNS'), ('for', 'IN'), ('unseen', 'JJ'), ('instances', 'NNS'), ('.', '.'), ('This', 'DT'), ('requires', 'VBZ'), ('the', 'DT'), ('learning', 'NN'), ('algorithm', 'NN'), ('to', 'TO'), ('generalize', 'VB'), ('from', 'IN'), ('the', 'DT'), ('training', 'NN'), ('data', 'NNS'), ('to', 'IN'), ('unseen', 'JJ'), ('situations', 'NNS'), ('in', 'IN'), ('a', 'DT'), ("'", '``'), ('reasonable', 'JJ'), ("'", "''"), ('way', 'NN'), ('(', ','), ('see', 'VB'), ('inductive', 'JJ'), ('bias', 'NN'), (')', '-RRB-'), ('.', '.'), ('Keywords', 'NNS'), ('are', 'VBP'), ('defined', 'VBN'), ('as', 'IN'), ('phrases', 'NNS'), ('that', 'WDT'), ('capture', 'VBP'), ('the', 'DT'), ('main', 'JJ'), ('topics', 'NNS'), ('discussed', 'VBN'), ('in', 'IN'), ('a', 'DT'), ('document', 'NN'), ('.', '.'), ('As', 'IN'), ('they', 'PRP'), ('offer', 'VBP'), ('a', 'DT'), ('brief', 'JJ'), ('yet', 'CC'), ('precise', 'JJ'), ('summary', 'NN'), ('of', 'IN'), ('document', 'NN'), ('content', 'NN'), (',', ','), ('they', 'PRP'), ('can', 'MD'), ('be', 'VB'), ('utilized', 'VBN'), ('for', 'IN'), ('various', 'JJ'), ('applications', 'NNS'), ('.', '.'), ('In', 'IN'), ('an', 'DT'), ('information', 'NN'), ('retrieval', 'NN'), ('environment', 'NN'), (',', ','), ('they', 'PRP'), ('serve', 'VBP'), ('as', 'IN'), ('an', 'DT'), ('indication', 'NN'), ('of', 'IN'), ('document', 'NN'), ('relevance', 'NN'), ('for', 'IN'), ('users', 'NNS'), (',', ','), ('as', 'IN'), ('the', 'DT'), ('list', 'NN'), ('of', 'IN'), ('keywords', 'NNS'), ('can', 'MD'), ('quickly', 'RB'), ('help', 'VB'), ('to', 'TO'), ('determine', 'VB'), ('whether', 'IN'), ('a', 'DT'), ('given', 'VBN'), ('document', 'NN'), ('is', 'VBZ'), ('relevant', 'JJ'), ('to', 'IN'), ('their', 'PRP$'), ('interest', 'NN'), ('.', '.'), ('As', 'IN'), ('keywords', 'NNS'), ('reflect', 'VBP'), ('a', 'DT'), ('document', 'NN'), ("'s", 'POS'), ('main', 'JJ'), ('topics', 'NNS'), (',', ','), ('they', 'PRP'), ('can', 'MD'), ('be', 'VB'), ('utilized', 'VBN'), ('to', 'TO'), ('classify', 'VB'), ('documents', 'NNS'), ('into', 'IN'), ('groups', 'NNS'), ('by', 'IN'), ('measuring', 'VBG'), ('the', 'DT'), ('overlap', 'NN'), ('between', 'IN'), ('the', 'DT'), ('keywords', 'NNS'), ('assigned', 'VBN'), ('to', 'IN'), ('them', 'PRP'), ('.', '.'), ('Keywords', 'NNS'), ('are', 'VBP'), ('also', 'RB'), ('used', 'VBN'), ('proactively', 'RB'), ('in', 'IN'), ('information', 'NN'), ('retrieval', 'NN'), ('.', '.')]
After the custom POS-tagger function is defined, it can be passed to KeyphraseVectorizers via the custom_pos_tagger parameter.
from keyphrase_vectorizers import KeyphraseCountVectorizer
# use custom POS-tagger with KeyphraseVectorizers
vectorizer = KeyphraseCountVectorizer(custom_pos_tagger=custom_pos_tagger)
vectorizer.fit(docs)
keyphrases = vectorizer.get_feature_names_out()
print(keyphrases)
>>> ['output value' 'information retrieval' 'algorithm' 'vector' 'groups'
'main topics' 'task' 'precise summary' 'supervised learning'
'inductive bias' 'information retrieval environment'
'supervised learning algorithm' 'function' 'input' 'pair'
'document relevance' 'learning' 'class labels' 'new examples' 'keywords'
'list' 'machine' 'training data' 'unseen situations' 'phrases' 'output'
'optimal scenario' 'document' 'training examples' 'documents' 'interest'
'indication' 'learning algorithm' 'inferred function'
'various applications' 'example' 'set' 'unseen instances'
'example input-output pairs' 'way' 'users' 'input object'
'supervisory signal' 'overlap' 'document content']
Using the keyphrase vectorizers together with KeyBERT for keyphrase extraction results in the PatternRank approach. PatternRank can extract grammatically correct keyphrases that are most similar to a document. Thereby, the vectorizer first extracts candidate keyphrases from the text documents, which are subsequently ranked by KeyBERT based on their document similarity. The top-n most similar keyphrases can then be considered as document keywords.
The advantage of using KeyphraseVectorizers in addition to KeyBERT is that it allows users to get grammatically correct
keyphrases instead of simple n-grams of pre-defined lengths. In KeyBERT, users can specify the keyphrase_ngram_range
to define the length of the retrieved keyphrases. However, this raises two issues. First, users usually do not know the
optimal n-gram range and therefore have to spend some time experimenting until they find a suitable n-gram range.
Second, even after finding a good n-gram range, the returned keyphrases are sometimes still grammatically not quite
correct or are slightly off-key. Unfortunately, this limits the quality of the returned keyphrases.
To adress this issue, we can use the vectorizers of this package to first extract candidate keyphrases that consist of
zero or more adjectives, followed by one or multiple nouns in a pre-processing step instead of simple n-grams. TextRank, SingleRank, and EmbedRank already successfully used this noun phrase approach for keyphrase extraction. The extracted candidate keyphrases are subsequently passed to KeyBERT for embedding generation and similarity calculation. To use both packages for keyphrase extraction, we need to
pass KeyBERT a keyphrase vectorizer with the vectorizer parameter. Since the length of keyphrases now depends on
part-of-speech tags, there is no need to define an n-gram length anymore.
KeyBERT can be installed via pip install keybert.
from keyphrase_vectorizers import KeyphraseCountVectorizer
from keybert import KeyBERT
docs = ["""Supervised learning is the machine learning task of learning a function that
maps an input to an output based on example input-output pairs. It infers a
function from labeled training data consisting of a set of training examples.
In supervised learning, each example is a pair consisting of an input object
(typically a vector) and a desired output value (also called the supervisory signal).
A supervised learning algorithm analyzes the training data and produces an inferred function,
which can be used for mapping new examples. An optimal scenario will allow for the
algorithm to correctly determine the class labels for unseen instances. This requires
the learning algorithm to generalize from the training data to unseen situations in a
'reasonable' way (see inductive bias).""",
"""Keywords are defined as phrases that capture the main topics discussed in a document.
As they offer a brief yet precise summary of document content, they can be utilized for various applications.
In an information retrieval environment, they serve as an indication of document relevance for users, as the list
of keywords can quickly help to determine whether a given document is relevant to their interest.
As keywords reflect a document's main topics, they can be utilized to classify documents into groups
by measuring the overlap between the keywords assigned to them. Keywords are also used proactively
in information retrieval."""]
kw_model = KeyBERT()
Instead of deciding on a suitable n-gram range which could be e.g.(1,2)...
>>> kw_model.extract_keywords(docs=docs, keyphrase_ngram_range=(1,2))
[[('labeled training', 0.6013),
('examples supervised', 0.6112),
('signal supervised', 0.6152),
('supervised', 0.6676),
('supervised learning', 0.6779)],
[('keywords assigned', 0.6354),
('keywords used', 0.6373),
('list keywords', 0.6375),
('keywords quickly', 0.6376),
('keywords defined', 0.6997)]]
we can now just let the keyphrase vectorizer decide on suitable keyphrases, without limitations to a maximum or minimum n-gram range. We only have to pass a keyphrase vectorizer as parameter to KeyBERT:
>>> kw_model.extract_keywords(docs=docs, vectorizer=KeyphraseCountVectorizer())
[[('learning', 0.4813),
('training data', 0.5271),
('learning algorithm', 0.5632),
('supervised learning', 0.6779),
('supervised learning algorithm', 0.6992)],
[('document content', 0.3988),
('information retrieval environment', 0.5166),
('information retrieval', 0.5792),
('keywords', 0.6046),
('document relevance', 0.633)]]
This allows us to make sure that we do not cut off important words caused by defining our n-gram range too short. For
example, we would not have found the keyphrase "supervised learning algorithm" with keyphrase_ngram_range=(1,2).
Furthermore, we avoid to get keyphrases that are slightly off-key like "labeled training", "signal supervised" or
"keywords quickly".
For more tips on how to use the KeyphraseVectorizers together with KeyBERT, visit this guide.
Similar to the application with KeyBERT, the keyphrase vectorizers can be used to obtain grammatically correct keyphrases as descriptions for topics instead of simple n-grams. This allows us to make sure that we do not cut off important topic description keyphrases by defining our n-gram range too short. Moreover, we don't need to clean stopwords upfront, can get more precise topic models and avoid to get topic description keyphrases that are slightly off-key.
BERTopic can be installed via pip install bertopic.
from keyphrase_vectorizers import KeyphraseCountVectorizer
from bertopic import BERTopic
from sklearn.datasets import fetch_20newsgroups
# load text documents
docs = fetch_20newsgroups(subset='all', remove=('headers', 'footers', 'quotes'))['data']
# only use subset of the data
docs = docs[:5000]
# train topic model with KeyphraseCountVectorizer
keyphrase_topic_model = BERTopic(vectorizer_model=KeyphraseCountVectorizer())
keyphrase_topics, keyphrase_probs = keyphrase_topic_model.fit_transform(docs)
# get topics
>>> keyphrase_topic_model.topics
{-1: [('file', 0.007265527630674131),
('one', 0.007055454904474792),
('use', 0.00633563957153475),
('program', 0.006053271092949018),
('get', 0.006011060091056076),
('people', 0.005729309058970368),
('know', 0.005635951168273583),
('like', 0.0055692449802916015),
('time', 0.00527028825803415),
('us', 0.00525564504880084)],
0: [('game', 0.024134589719090525),
('team', 0.021852806383170772),
('players', 0.01749406934044139),
('games', 0.014397938026886745),
('hockey', 0.013932342023677305),
('win', 0.013706115572901401),
('year', 0.013297593024390321),
('play', 0.012533185558169046),
('baseball', 0.012412743802062559),
('season', 0.011602725885164318)],
1: [('patients', 0.022600352291162015),
('msg', 0.02023877371575874),
('doctor', 0.018816282737587457),
('medical', 0.018614407917995103),
('treatment', 0.0165028251400717),
('food', 0.01604980195180696),
('candida', 0.015255961242066143),
('disease', 0.015115496310099693),
('pain', 0.014129703072484495),
('hiv', 0.012884503220341102)],
2: [('key', 0.028851633177510126),
('encryption', 0.024375137861044675),
('clipper', 0.023565947302544528),
('privacy', 0.019258719348097385),
('security', 0.018983682856076434),
('chip', 0.018822199098878365),
('keys', 0.016060139239615384),
('internet', 0.01450486904722165),
('encrypted', 0.013194373119964168),
('government', 0.01303978311708837)],
...
The same topics look a bit different when no keyphrase vectorizer is used:
from bertopic import BERTopic
from sklearn.datasets import fetch_20newsgroups
# load text documents
docs = fetch_20newsgroups(subset='all', remove=('headers', 'footers', 'quotes'))['data']
# only use subset of the data
docs = docs[:5000]
# train topic model without KeyphraseCountVectorizer
topic_model = BERTopic()
topics, probs = topic_model.fit_transform(docs)
# get topics
>>> topic_model.topics
{-1: [('the', 0.012864641020408933),
('to', 0.01187920529994724),
('and', 0.011431498631699856),
('of', 0.01099851927541331),
('is', 0.010995478673036962),
('in', 0.009908233622158523),
('for', 0.009903667215879675),
('that', 0.009619596716087699),
('it', 0.009578499681829809),
('you', 0.0095328846440753)],
0: [('game', 0.013949166096523719),
('team', 0.012458483177116456),
('he', 0.012354733462693834),
('the', 0.01119583508278812),
('10', 0.010190243555226108),
('in', 0.0101436249231417),
('players', 0.009682212470082758),
('to', 0.00933700544705287),
('was', 0.009172402203816335),
('and', 0.008653375901739337)],
1: [('of', 0.012771267188340924),
('to', 0.012581337590513296),
('is', 0.012554884458779008),
('patients', 0.011983273578628046),
('and', 0.011863499662237566),
('that', 0.011616113472989725),
('it', 0.011581944987387165),
('the', 0.011475148304229873),
('in', 0.011395485985801054),
('msg', 0.010715000656335596)],
2: [('key', 0.01725282988290282),
('the', 0.014634841495851404),
('be', 0.014429762197907552),
('encryption', 0.013530733999898166),
('to', 0.013443159534369817),
('clipper', 0.01296614319927958),
('of', 0.012164734232650158),
('is', 0.012128295958613464),
('and', 0.011972763728732667),
('chip', 0.010785744492767285)],
...
The KeyphraseVectorizers also support online/incremental updates of their representation (similar to the OnlineCountVectorizer). The vectorizer can not only update out-of-vocabulary keyphrases but also implements decay and cleaning functions to prevent the sparse document-keyphrases matrix to become too large.
Parameters for online updates:
decay: At each iteration, we sum the document-keyphrase representation of the new documents with the
document-keyphrase representation of all documents processed thus far. In other words, the document-keyphrase matrix
keeps increasing with each iteration. However, especially in a streaming setting, older documents might become less
and less relevant as time goes on. Therefore, a decay parameter was implemented that decays the document-keyphrase
frequencies at each iteration before adding the document frequencies of new documents. The decay parameter is a value
between 0 and 1 and indicates the percentage of frequencies the previous document-keyphrase matrix should be reduced
to. For example, a value of .1 will decrease the frequencies in the document-keyphrase matrix by 10% at each iteration
before adding the new document-keyphrase matrix. This will make sure that recent data has more weight than previous
iterations.delete_min_df: We might want to remove keyphrases from the document-keyphrase representation that appear
infrequently. The min_df parameter works quite well for that. However, when we have a streaming setting,
the min_df does not work as well since a keyphrases's frequency might start below min_df but will end up higher
than that over time. Setting that value high might not always be advised. As a result, the list of keyphrases learned
by the vectorizer and the resulting document-keyphrase matrix can become quite large. Similarly, if we implement
the decay parameter, then some values will decrease over time until they are below min_df. For these reasons,
the delete_min_df parameter was implemented. The parameter takes positive integers and indicates, at each iteration,
which keyphrases will be removed from the already learned ones. If the value is set to 5, it will check after each
iteration if the total frequency of a keyphrase is exceeded by that value. If so, the keyphrase will be removed in its
entirety from the list of keyphrases learned by the vectorizer. This helps to keep the document-keyphrase matrix of a
manageable size.from keyphrase_vectorizers import KeyphraseCountVectorizer
docs = ["""Supervised learning is the machine learning task of learning a function that
maps an input to an output based on example input-output pairs. It infers a
function from labeled training data consisting of a set of training examples.
In supervised learning, each example is a pair consisting of an input object
(typically a vector) and a desired output value (also called the supervisory signal).
A supervised learning algorithm analyzes the training data and produces an inferred function,
which can be used for mapping new examples. An optimal scenario will allow for the
algorithm to correctly determine the class labels for unseen instances. This requires
the learning algorithm to generalize from the training data to unseen situations in a
'reasonable' way (see inductive bias).""",
"""Keywords are defined as phrases that capture the main topics discussed in a document.
As they offer a brief yet precise summary of document content, they can be utilized for various applications.
In an information retrieval environment, they serve as an indication of document relevance for users, as the list
of keywords can quickly help to determine whether a given document is relevant to their interest.
As keywords reflect a document's main topics, they can be utilized to classify documents into groups
by measuring the overlap between the keywords assigned to them. Keywords are also used proactively
in information retrieval."""]
# Init default vectorizer.
vectorizer = KeyphraseCountVectorizer(decay=0.5, delete_min_df=3)
# intitial vectorizer fit
vectorizer.fit_transform([docs[0]]).toarray()
>>> array([[1, 1, 3, 1, 1, 3, 1, 3, 1, 1, 1, 1, 2, 1, 3, 1, 1, 1, 1, 3, 1, 3,
1, 1, 1]])
# check learned keyphrases
print(vectorizer.get_feature_names_out())
>>> ['output pairs', 'output value', 'function', 'optimal scenario',
'pair', 'supervised learning', 'supervisory signal', 'algorithm',
'supervised learning algorithm', 'way', 'training examples',
'input object', 'example', 'machine', 'output',
'unseen situations', 'unseen instances', 'inductive bias',
'new examples', 'input', 'task', 'training data', 'class labels',
'set', 'vector']
# learn additional keyphrases from new documents with partial fit
vectorizer.partial_fit([docs[1]])
vectorizer.transform([docs[1]]).toarray()
>>> array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 1, 1, 2, 1, 1, 2, 1, 1, 1, 1, 1, 1, 5, 1, 1, 5, 1]])
# check learned keyphrases, including newly learned ones
print(vectorizer.get_feature_names_out())
>>> ['output pairs', 'output value', 'function', 'optimal scenario',
'pair', 'supervised learning', 'supervisory signal', 'algorithm',
'supervised learning algorithm', 'way', 'training examples',
'input object', 'example', 'machine', 'output',
'unseen situations', 'unseen instances', 'inductive bias',
'new examples', 'input', 'task', 'training data', 'class labels',
'set', 'vector', 'list', 'various applications',
'information retrieval', 'groups', 'overlap', 'main topics',
'precise summary', 'document relevance', 'interest', 'indication',
'information retrieval environment', 'phrases', 'keywords',
'document content', 'documents', 'document', 'users']
# update list of learned keyphrases according to 'delete_min_df'
vectorizer.update_bow([docs[1]])
vectorizer.transform([docs[1]]).toarray()
>>> array([[5, 5]])
# check updated list of learned keyphrases (only the ones that appear more than 'delete_min_df' remain)
print(vectorizer.get_feature_names_out())
>>> ['keywords', 'document']
# update again and check the impact of 'decay' on the learned document-keyphrase matrix
vectorizer.update_bow([docs[1]])
vectorizer.X_.toarray()
>>> array([[7.5, 7.5]])
When citing KeyphraseVectorizers or PatternRank in academic papers and theses, please use this BibTeX entry:
@conference{schopf_etal_kdir22,
author={Tim Schopf and Simon Klimek and Florian Matthes},
title={PatternRank: Leveraging Pretrained Language Models and Part of Speech for Unsupervised Keyphrase Extraction},
booktitle={Proceedings of the 14th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management - KDIR},
year={2022},
pages={243-248},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0011546600003335},
isbn={978-989-758-614-9},
issn={2184-3228},
}
FAQs
Set of vectorizers that extract keyphrases with part-of-speech patterns from a collection of text documents and convert them into a document-keyphrase matrix.
We found that keyphrase-vectorizers demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.
Did you know?
Socket for GitHub automatically highlights issues in each pull request and monitors the health of all your open source dependencies. Discover the contents of your packages and block harmful activity before you install or update your dependencies.
Security News
Research
The Socket Research Team breaks down a malicious wrapper package that uses obfuscation to harvest credentials and exfiltrate sensitive data.
Research
Security News
Attackers used a malicious npm package typosquatting a popular ESLint plugin to steal sensitive data, execute commands, and exploit developer systems.
Security News
The Ultralytics' PyPI Package was compromised four times in one weekend through GitHub Actions cache poisoning and failure to rotate previously compromised API tokens.