textacy 0.2.5

Higher-level text processing, built on Spacy

textacy is a Python library for performing higher-level natural language processing (NLP) tasks, built on the high-performance spaCy library. With the basics — tokenization, part-of-speech tagging, parsing — offloaded to another library, textacy focuses on tasks facilitated by the availability of tokenized, POS-tagged, and parsed text: keyterm extraction, readability statistics, emotional valence analysis, quotation attribution, and more.

Features

• Functions for preprocessing raw text prior to analysis (whitespace normalization, URL/email/number/date replacement, unicode fixing/stripping, etc.)

• Convenient interface to basic linguistic elements provided by Spacy (words, ngrams, noun phrases, etc.), along with standardized filtering options

• Variety of functions for extracting information from text (particular POS patterns, subject-verb-object triples, acronyms and their definitions, direct quotations, etc.)

• Unsupervised key term extraction (specific algorithms such as SGRank or TextRank, as well as a general semantic network-based approach)

• Conversion of individual documents into common representations (bag of words), as well as corpora (term-document matrix, with TF or TF-IDF weighting, and filtering by these metrics or IC)

• Common utility functions for identifying a text’s language, displaying key words in context (KWIC), truecasing words, and higher-level navigation of a parse tree

• Sklearn-style topic modeling with LSA, LDA, or NMF, including functions to interpret the results of trained models

And more!

Installation

The simple way to install textacy is

$ pip install -U textacy

Or, download and unzip the source tar.gz from PyPi, then

$ python setup.py install

Example

>>> import textacy

Efficiently stream documents from disk and into a processed corpus:

>>> docs = textacy.corpora.fetch_bernie_and_hillary()
>>> content_stream, metadata_stream = textacy.fileio.split_content_and_metadata(
...     docs, 'text', itemwise=False)
>>> corpus = textacy.TextCorpus.from_texts(
...     'en', content_stream, metadata_stream, n_threads=2)
>>> print(corpus)
TextCorpus(3066 docs; 1909705 tokens)

Represent corpus as a document-term matrix, with flexible weighting and filtering:

>>> doc_term_matrix, id2term = corpus.as_doc_term_matrix(
...     (doc.as_terms_list(words=True, ngrams=False, named_entities=True)
...      for doc in corpus),
...     weighting='tfidf', normalize=True, smooth_idf=True, min_df=2, max_df=0.95)
>>> print(repr(doc_term_matrix))
<3066x16145 sparse matrix of type '<class 'numpy.float64'>'
    with 432067 stored elements in Compressed Sparse Row format>

Train and interpret a topic model:

>>> model = textacy.tm.TopicModel('nmf', n_topics=10)
>>> model.fit(doc_term_matrix)
>>> doc_topic_matrix = model.transform(doc_term_matrix)
>>> print(doc_topic_matrix.shape)
(3066, 10)
>>> for topic_idx, top_terms in model.top_topic_terms(id2term, top_n=10):
...     print('topic', topic_idx, ':', '   '.join(top_terms))
topic 0 : people   tax   $   percent   american   million   republican   country   go   americans
topic 1 : rescind   quorum   order   consent   unanimous   ask   president   mr.   madam   absence
topic 2 : chairman   chairman.   amendment   mr.   clerk   gentleman   designate   offer   sanders   vermont
topic 3 : dispense   reading   amendment   consent   unanimous   ask   president   mr.   madam   pending
topic 4 : senate   consent   session   unanimous   authorize   ask   committee   meet   president   a.m.
topic 5 : health   care   state   child   veteran   va   vermont   new   's   need
topic 6 : china   american   speaker   worker   trade   job   wage   america   gentleman   people
topic 7 : social security   social   security   cut   senior   medicare   deficit   benefit   year   cola
topic 8 : senators   desiring   chamber   vote   minute   morning   permit   10 minute   proceed   speak
topic 9 : motion   table   reconsider   lay   agree   preamble   record   resolution   consent   print

Basic indexing as well as flexible selection of documents in a corpus:

>>> bernie_docs = list(corpus.get_docs(
...     lambda doc: doc.metadata['speaker'] == 'Bernard Sanders'))
>>> print(len(bernie_docs))
2236
>>> doc = corpus[-1]
>>> print(doc)
TextDoc(465 tokens; "Mr. President, I ask to have printed in the Rec...")

Preprocess plain text, or highlight particular terms in it:

>>> textacy.preprocess_text(doc.text, lowercase=True, no_punct=True)[:70]
'mr president i ask to have printed in the record copies of some of the'
>>> textacy.text_utils.keyword_in_context(doc.text, 'nation', window_width=35)
ed States of America is an amazing  nation  that continues to lead the world t
come the role model for developing  nation s attempting to give their people t
ve before to better ourselves as a  nation , because what we change will set a
nd education. Fortunately, we as a  nation  have the opportunity to fix the in
 sentences. Judges from across the  nation  have said for decades that they do
reopened many racial wounds in our  nation . The war on drugs also put addicts

Extract various elements of interest from parsed documents:

>>> list(doc.ngrams(2, filter_stops=True, filter_punct=True, filter_nums=False))[:15]
[Mr. President,
 Record copies,
 finalist essays,
 essays written,
 Vermont High,
 High School,
 School students,
 sixth annual,
 annual ``,
 essay contest,
 contest conducted,
 nearly 800,
 800 entries,
 material follows,
 United States]
>>> list(doc.ngrams(3, filter_stops=True, filter_punct=True, min_freq=2))
[lead the world,
 leading the world,
 2.2 million people,
 2.2 million people,
 mandatory minimum sentences,
 Mandatory minimum sentences,
 war on drugs,
 war on drugs]
>>> list(doc.named_entities(drop_determiners=True, bad_ne_types='numeric'))
[Record,
 Vermont High School,
 United States of America,
 Americans,
 U.S.,
 U.S.,
 African American]
>>> pattern = textacy.regexes_etc.POS_REGEX_PATTERNS['en']['NP']
>>> print(pattern)
<DET>? <NUM>* (<ADJ> <PUNCT>? <CONJ>?)* (<NOUN>|<PROPN> <PART>?)+
>>> list(doc.pos_regex_matches(pattern))[-10:]
[experiment,
 many racial wounds,
 our nation,
 The war,
 drugs,
 addicts,
 bars,
 addiction,
 the problem,
 a mental health issue]
>>> list(doc.semistructured_statements('it', cue='be'))
[(it, is, important to humanize these statistics),
 (It, is, the third highest state expenditure, behind health care and education),
 (it, is, ; a mental health issue)]
>>> doc.key_terms(algorithm='textrank', n=5)
[('nation', 0.04315758994993049),
 ('world', 0.030590559641614556),
 ('incarceration', 0.029577233127175532),
 ('problem', 0.02411902162606202),
 ('people', 0.022631145896105508)]

Compute common statistical attributes of a text:

>>> doc.readability_stats
{'automated_readability_index': 11.67580188679245,
 'coleman_liau_index': 10.89927271226415,
 'flesch_kincaid_grade_level': 10.711962264150948,
 'flesch_readability_ease': 56.022660377358505,
 'gunning_fog_index': 13.857358490566037,
 'n_chars': 2026,
 'n_polysyllable_words': 57,
 'n_sents': 20,
 'n_syllables': 648,
 'n_unique_words': 228,
 'n_words': 424,
 'smog_index': 12.773325707644965}

Count terms individually, and represent documents as a bag of terms with flexible weighting and inclusion criteria:

>>> doc.term_count('nation')
6
>>> bot = doc.as_bag_of_terms(weighting='tf', normalized=False, lemmatize='auto', ngram_range=(1, 1))
>>> [(doc.spacy_stringstore[term_id], count)
...  for term_id, count in bot.most_common(n=10)]
[('nation', 6),
 ('world', 4),
 ('incarceration', 4),
 ('people', 3),
 ('mandatory minimum', 3),
 ('lead', 3),
 ('minimum', 3),
 ('problem', 3),
 ('mandatory', 3),
 ('drug', 3)]

Project Links

• textacy @ PyPi

• textacy @ GitHub

• textacy @ ReadTheDocs

Authors

• Burton DeWilde (<[e-mail]>)

Unofficial Roadmap

• [x] import/export for common formats

• [x] serialization and streaming to/from disk

• [x] topic modeling via gensim and/or sklearn

• [x] data viz for text analysis

• [ ] distributional representations (word2vec etc.) via either gensim or spacy

• [ ] document similarity/clustering (?)

• [ ] basic dictionary-based methods e.g. sentiment analysis (?)

• [ ] text classification

• [ ] media frames analysis