Simple Text Analysis Using Python – Identifying Named Entities, Tagging, Fuzzy String Matching and Topic Modelling

Text processing is not really my thing, but here’s a round-up of some basic recipes that allow you to get started with some quick’n’dirty tricks for identifying named entities in a document, and tagging entities in documents.

In this post, I’ll briefly review some getting started code for:

  • performing simple entity extraction from a text; for example, when presented with a text document, label it with named entities (people, places, organisations); entity extraction is typically based on statistical models that rely on document features such as correct capitalisation of names to work correctly;
  • tagging documents that contain exact matches of specified terms: in this case, we have a list of specific text strings we are interested in (for example, names of people or companies) and we want to know if there are exact matches in the text document and where those matches occur in the document;
  • partial and fuzzing string matching of specified entities in a text: in this case, we may want to know whether something resembling a specified text string occurs in the document (for example, mis0spellings of name);
  • topic modelling: the identification, using statistical models, of “topic terms” that appear across a set of documents.

You can find a gist containing a notebook that summarises the code here.

Simple named entity recognition

spaCy is a natural language processing library for Python library that includes a basic model capable of recognising (ish!) names of people, places and organisations, as well as dates and financial amounts.

According to the spaCy entity recognition documentation, the built in model recognises the following types of entity:

  • PERSON People, including fictional.
  • NORP Nationalities or religious or political groups.
  • FACILITY Buildings, airports, highways, bridges, etc.
  • ORG Companies, agencies, institutions, etc.
  • GPE Countries, cities, states. (That is, Geo-Political Entitites)
  • LOC Non-GPE locations, mountain ranges, bodies of water.
  • PRODUCT Objects, vehicles, foods, etc. (Not services.)
  • EVENT Named hurricanes, battles, wars, sports events, etc.
  • WORK_OF_ART Titles of books, songs, etc.
  • LANGUAGE Any named language.
  • LAW A legislation related entity(?)

Quantities are also recognised:

  • DATE Absolute or relative dates or periods.
  • TIME Times smaller than a day.
  • PERCENT Percentage, including “%”.
  • MONEY Monetary values, including unit.
  • QUANTITY Measurements, as of weight or distance.
  • ORDINAL “first”, “second”, etc.
  • CARDINAL Numerals that do not fall under another type.

Custom models can also be trained, but this requires annotated training documents.

#!pip3 install spacy
from spacy.en import English
parser = English()
That this House notes the announcement of 300 redundancies at the Nestlé manufacturing factories in York, Fawdon, Halifax and Girvan and that production of the Blue Riband bar will be transferred to Poland; acknowledges in the first three months of 2017 Nestlé achieved £21 billion in sales, a 0.4 per cent increase over the same period in 2016; further notes 156 of these job losses will be in York, a city that in the last six months has seen 2,000 job losses announced and has become the most inequitable city outside of the South East, and a further 110 jobs from Fawdon, Newcastle; recognises the losses come within a month of triggering Article 50, and as negotiations with the EU on the UK leaving the EU and the UK's future with the EU are commencing; further recognises the cost of importing products, including sugar, cocoa and production machinery, has risen due to the weakness of the pound and the uncertainty over the UK's future relationship with the single market and customs union; and calls on the Government to intervene and work with hon. Members, trades unions GMB and Unite and the company to avert these job losses now and prevent further job losses across Nestlé.
#Code "borrowed" from somewhere?!
def entities(example, show=False):
    if show: print(example)
    parsedEx = parser(example)

    print("-------------- entities only ---------------")
    # if you just want the entities and nothing else, you can do access the parsed examples "ents" property like this:
    ents = list(parsedEx.ents)
    for entity in ents:
        #print(entity.label, entity.label_, ' '.join(t.orth_ for t in entity))
        term=' '.join(t.orth_ for t in entity)
        if ' '.join(term) not in tags:
            tags[term]=[(entity.label, entity.label_)]
            tags[term].append((entity.label, entity.label_))
-------------- entities only ---------------
{'House': [(380, 'ORG')], '300': [(393, 'CARDINAL')], 'Nestlé': [(380, 'ORG')], '\n York , Fawdon': [(381, 'GPE')], 'Halifax': [(381, 'GPE')], 'Girvan': [(381, 'GPE')], 'the Blue Riband': [(380, 'ORG')], 'Poland': [(381, 'GPE')], '\n': [(381, 'GPE'), (381, 'GPE')], 'the first three months of 2017': [(387, 'DATE')], '£ 21 billion': [(390, 'MONEY')], '0.4 per': [(390, 'MONEY')], 'the same period in 2016': [(387, 'DATE')], '156': [(393, 'CARDINAL')], 'York': [(381, 'GPE')], '\n the': [(381, 'GPE')], 'six': [(393, 'CARDINAL')], '2,000': [(393, 'CARDINAL')], 'the South East': [(382, 'LOC')], '110': [(393, 'CARDINAL')], 'Fawdon': [(381, 'GPE')], 'Newcastle': [(380, 'ORG')], 'a month of': [(387, 'DATE')], 'Article 50': [(21153, 'LAW')], 'EU': [(380, 'ORG')], 'UK': [(381, 'GPE')], 'GMB': [(380, 'ORG')], 'Unite': [(381, 'GPE')]}
q= "Bob Smith was in the Houses of Parliament the other day"
-------------- entities only ---------------
{'Bob Smith': [(377, 'PERSON')]}

Note that the way that models are trained typically realises on cues from the correct capitalisation of named entities.

-------------- entities only ---------------


A simplistic, and quite slow, tagger, supporting limited recognition of Locations (I-LOC), Organizations (I-ORG) and Persons (I-PER).

#!pip3 install polyglot

##Mac ??
#!brew install icu4c
#I found I needed: pip3 install pyicu, pycld2, morfessor
#apt-get install libicu-dev
!polyglot download embeddings2.en ner2.en
[polyglot_data] Downloading package embeddings2.en to
[polyglot_data]     /Users/ajh59/polyglot_data...
[polyglot_data] Downloading package ner2.en to
[polyglot_data]     /Users/ajh59/polyglot_data...
from polyglot.text import Text

text = Text(example)
[I-PER(['Bob', 'Smith'])]

Partial Matching Specific Entities

Sometimes we may have a list of entities that we wish to match in a text. For example, suppose we have a list of MPs’ names, or a list of ogranisations of subject terms identified in a thesaurus, and we want to tag a set of documents with those entities if the entity exists in the document.

To do this, we can search a text for strings that exactly match any of the specified terms or where any of the specified terms match part of a longer string in the text.

Naive implementations can take a signifcant time to find multiple strings within a tact, but the Aho-Corasick algorithm will efficiently match a large set of key values within a particular text.

## The following recipe was hinted at via @pudo

#!pip3 install pyahocorasick

First, construct an automaton that identifies the terms you want to detect in the target text.

from ahocorasick import Automaton

A.add_word("European Union",('VOCAB','European Union'))
A.add_word("Boris Johnson",('PERSON','Boris Johnson'))
A.add_word("Boris",('PERSON','Boris Johnson'))
A.add_word("Boris Johnson",('PERSON','Boris Johnson (LC)'))

q2='Boris Johnson went off to Europe to complain about the European Union'
for item in A.iter(q2):
    print(item, q2[:item[0]+1])
(4, ('PERSON', 'Boris Johnson')) Boris
(12, ('PERSON', 'Boris Johnson')) Boris Johnson
(31, ('VOCAB', 'Europe')) Boris Johnson went off to Europe
(60, ('VOCAB', 'Europe')) Boris Johnson went off to Europe to complain about the Europe
(68, ('VOCAB', 'European Union')) Boris Johnson went off to Europe to complain about the European Union

Once again, case is important.

q2l = q2.lower()
for item in A.iter(q2l):
    print(item, q2l[:item[0]+1])
(12, ('PERSON', 'Boris Johnson (LC)')) boris johnson

We can tweak the automata patterns to capture the length of the string match term, so we can annotate the text with matches more exactly:

A.add_word("Europe",(('VOCAB', len("Europe")),'Europe'))
A.add_word("European Union",(('VOCAB', len("European Union")),'European Union'))
A.add_word("Boris Johnson",(('PERSON', len("Boris Johnson")),'Boris Johnson'))
A.add_word("Boris",(('PERSON', len("Boris")),'Boris Johnson'))

for item in A.iter(q2):
    print(item, '{}*{}*{}'.format(q2[start-3:start],q2[start:end],q2[end:end+3]))
(4, (('PERSON', 5), 'Boris Johnson')) *Boris* Jo
(12, (('PERSON', 13), 'Boris Johnson')) *Boris Johnson* we
(31, (('VOCAB', 6), 'Europe')) to *Europe* to
(60, (('VOCAB', 6), 'Europe')) he *Europe*an 
(68, (('VOCAB', 14), 'European Union')) he *European Union*

Fuzzy String Matching

Whilst the Aho-Corasick approach will return hits for strings in the text that partially match the exact match key terms, sometimes we want to know whether there are terms in a text that almost match terms in specific set of terms.

Imagine a situation where we have managed to extract arbitrary named entities from a text, but they do not match strings in a specified list in an exact or partially exact way. Our next step might be to attempt to further match those entities in a fuzzy way with entities in a specified list.


The python fuzzyset package will try to match a specified string to similar strings in a list of target strings, returning a single item from a specified target list that best matches the provided term.

For example, if we extract the name Boris Johnstone in a text, we might then try to further match that string, in a fuzzy way, with a list of correctly spelled MP names.

A confidence value expresses the degree of match to terms in the fuzzy match set list.

import fuzzyset

fz = fuzzyset.FuzzySet()
#Create a list of terms we would like to match against in a fuzzy way
for l in ["Diane Abbott", "Boris Johnson"]:

#Now see if our sample term fuzzy matches any of those specified terms
sample_term='Boris Johnstone'
fz.get(sample_term), fz.get('Diana Abbot'), fz.get('Joanna Lumley')
([(0.8666666666666667, 'Boris Johnson')],
 [(0.8333333333333334, 'Diane Abbott')],
 [(0.23076923076923073, 'Diane Abbott')])


If we want to try to find a fuzzy match for a term within a text, we can use the python fuzzywuzzy library. Once again, we spcify a list of target items we want to try to match against.

from fuzzywuzzy import process
from fuzzywuzzy import fuzz
terms=['Houses of Parliament', 'Diane Abbott', 'Boris Johnson']

q= "Diane Abbott, Theresa May and Boris Johnstone were in the Houses of Parliament the other day"
[('Houses of Parliament', 90), ('Diane Abbott', 90), ('Boris Johnson', 86)]

By default, we get match confidence levels for each term in the target match set, although we can limit the response to a maximum number of matches:

process.extract(q,terms,scorer=fuzz.partial_ratio, limit=2)
[('Houses of Parliament', 90), ('Boris Johnson', 85)]

A range of fuzzy match scroing algorithms are supported:

  • WRatio – measure of the sequences’ similarity between 0 and 100, using different algorithms
  • QRatio – Quick ratio comparison between two strings
  • UWRatio – a measure of the sequences’ similarity between 0 and 100, using different algorithms. Same as WRatio but preserving unicode
  • UQRatio – Unicode quick ratio
  • ratio
  • `partial_ratio – ratio of the most similar substring as a number between 0 and 100
  • token_sort_ratio – a measure of the sequences’ similarity between 0 and 100 but sorting the token before comparing
  • partial_token_set_ratio
  • partial_token_sort_ratio – ratio of the most similar substring as a number between 0 and 100 but sorting the token before comparing

More usefully, perhaps, is to return items that match above a particular confidence level:

[('Houses of Parliament', 90), ('Diane Abbott', 90)]

However, one problem with the fuzzywuzzy matcher is that it doesn’t tell us where in the supplied text string the match occurred, or what string in the text was matched.

The fuzzywuzzy package can also be used to try to deduplicate a list of items, returning the longest item in the duplicate list. (It might be more useful if this is optionally the first item in the original list?)

names=['Diane Abbott', 'Boris Johnson','Boris Johnstone','Diana Abbot', 'Boris Johnston','Joanna Lumley']
process.dedupe(names, threshold=80)
['Joanna Lumley', 'Boris Johnstone', 'Diane Abbott']

It might also be useful to see the candidate strings associated with each deduped item, treating the first item in the list as the canonical one:

import hashlib

for t in names:
    #Generate a key based on the sorted members of the set
    keyvals=sorted(set([x[0] for x in fuzzyset]),key=lambda x:names.index(x),reverse=False)
    if len(keyvals)>1 and key not in fuzzed:
        clusters[key]=sorted(set([x for x in fuzzyset]),key=lambda x:names.index(x[0]),reverse=False)
for cluster in clusters:
[('Diane Abbott', 100), ('Diana Abbot', 87)]
[('Boris Johnson', 100), ('Boris Johnstone', 93), ('Boris Johnston', 96)]

OpenRefine Clustering

As well as running as a browser accessed application, OpenRefine also runs as a service that can be accessed from Python using the client libary.

In particular, we can use the OpenRefine service to cluster fuzzily matched items within a list of items.

#!pip install git+
#NOTE - this requires a python 2 kernel
#Initialise the connection to the server using default or environment variable defined server settings
#REFINE_HOST = os.environ.get('OPENREFINE_HOST', os.environ.get('GOOGLE_REFINE_HOST', ''))
#REFINE_PORT = os.environ.get('OPENREFINE_PORT', os.environ.get('GOOGLE_REFINE_PORT', '3333'))
from google.refine import refine, facet
server = refine.RefineServer()
orefine = refine.Refine(server)
#Create an example CSV file to load into a test OpenRefine project
project_file = 'simpledemo.csv'
with open(project_file,'w') as f:
    for t in ['Name']+names+['Boris Johnstone']:
        f.write(t+ '\n')
!cat {project_file}
Diane Abbott
Boris Johnson
Boris Johnstone
Diana Abbot
Boris Johnston
Joanna Lumley
Boris Johnstone

OpenRefine supports a range of clustering functions:

- clusterer_type: binning; function: fingerprint|metaphone3|cologne-phonetic
- clusterer_type: binning; function: ngram-fingerprint; params: {'ngram-size': INT}
- clusterer_type: knn; function: levenshtein|ppm; params: {'radius': FLOAT,'blocking-ngram-size': INT}
for cluster in clusters:
[{'count': 1, 'value': u'Diana Abbot'}, {'count': 1, 'value': u'Diane Abbott'}]
[{'count': 2, 'value': u'Boris Johnstone'}, {'count': 1, 'value': u'Boris Johnston'}]

Topic Models

Topic models are statistical models that attempts to categorise different “topics” that occur across a set of docments.

Several python libraries provide a simple interface for the generation of topic models from text contained in multiple documents.


#!pip3 install gensim
from gensim import corpora, models

def get_lda_from_lists_of_words(lists_of_words, **kwargs):
    dictionary = corpora.Dictionary(lists_of_words) # this dictionary maps terms to integers
    corpus = [dictionary.doc2bow(text) for text in lists_of_words] # create a bag of words from each document
    tfidf = models.TfidfModel(corpus) # this models the significance of words using term frequency inverse document frequency
    corpus_tfidf = tfidf[corpus]
    kwargs["id2word"] = dictionary # set the dictionary
    return models.LdaModel(corpus_tfidf, **kwargs) # do the LDA topic modelling

def print_top_terms(lda, num_terms=10):
    for i in range(0, num_terms):
        terms = [term for term,val in lda.show_topic(i,num_terms)]
        txt.append("\t - top {} terms for topic #{}: {}".format(num_terms,i,' '.join(terms)))
    return '\n'.join(txt)

To start with, let’s create a list of dummy documents and then generate word lists for each document.

docs=['The banks still have a lot to answer for the financial crisis.',
     'This MP and that Member of Parliament were both active in the debate.',
     'The companies that work in finance need to be responsible.',
     'There is a reponsibility incumber on all participants for high quality debate in Parliament.',
     'Corporate finance is a big responsibility.']

#Create lists of words from the text in each document
from nltk.tokenize import word_tokenize
docs = [ word_tokenize(doc.lower()) for doc in docs ]

#Remove stop words from the wordlists
from nltk.corpus import stopwords
docs = [ [word for word in doc if word not in stopwords.words('english') ] for doc in docs ]

Now we can generate the topic models from the list of word lists.

topicsLda = get_lda_from_lists_of_words([s for s in docs if isinstance(s,list)], num_topics=3, passes=20)
print( print_top_terms(topicsLda))
     - top 3 terms for topic #0: parliament debate active
     - top 3 terms for topic #1: responsible work need
     - top 3 terms for topic #2: corporate big responsibility

The model is randomised – if we run it again we are likely to get a different result.

topicsLda = get_lda_from_lists_of_words([s for s in docs if isinstance(s,list)], num_topics=3, passes=20)
print( print_top_terms(topicsLda))
     - top 3 terms for topic #0: finance corporate responsibility
     - top 3 terms for topic #1: participants quality high
     - top 3 terms for topic #2: member mp active

So What Can Text Analysis Do for You?

Despite believing we can treat anything we can represent in digital form as “data”, I’m still pretty flakey on understanding what sorts of analysis we can easily do with different sorts of data. Time series analysis is one area – the pandas Python library has all manner of handy tools for working with that sort of data that I have no idea how to drive – and text analysis is another.

So prompted by Sheila MacNeill’s post about textexture, which I guessed might be something to do with topic modeling (I should have read the about, h/t @mhawksey), here’s a quick round up of handy things the text analysts seem to be able to do pretty easily…

Taking the lazy approach, I has a quick look at the CRAN natural language processing task view to get an idea of what sort of tool support for text analysis there is in R, and a peek through the NLTK documentation to see what sort of thing we might be readily able to do in Python. Note that this take is a personal one, identifying the sorts of things that I can see I might personally have a recurring use for…

First up – extracting text from different document formats. I’ve already posted about Apache Tika, which can pull text from a wide range of documents (PDFs, extract text from Word docs, extract text from images), which seems to be a handy, general purpose tool. (Other tools are available, but I only have so much time, and for now Tika seems to do what I need…)

Second up, concordance views. The NLTK docs describe concordance views as follows: “A concordance view shows us every occurrence of a given word, together with some context.” So for example:


This can be handy for skimming through multiple references to a particular item, rather than having to do a lot of clicking, scrolling or page turning.

How about if we want to compare the near co-occurrence of words or phrases in a document? One way to do this is graphically, plotting the “distance” through the text on the x-axis, and then for categorical terms on y marking out where those terms appear in the text. In NLTK, this is referred to as a lexical dispersion plot:


I guess we could then scan across the distance axis using a windowing function to find terms that appear within a particular distance of each other? Or use co-occurrence matrices for example (eg Co-occurrence matrices of time series applied to literary works), perhaps with overlapping “time” bins? (This could work really well as a graph model – eg for 20 pages, set up page nodes 1-2, 2-3, 3-4,.., 18-19, 19-20, then an actor node for each actor, connecting actors to page nodes for page bins on which they occur; then project the bipartite graph onto just the actor nodes, connecting actors who were originally to the same page bin nodes.)

Something that could be differently useful is spotting common sentences that appear in different documents (for example, quotations). There are surely tools out there that do this, though offhand I can’t find any..? My gut reaction would be to generate a sentence list for each document (eg using something like the handy looking textblob python library), strip quotation marks and whitespace, etc, sort each list, then run a diff on them and pull out the matched lines. (So a “reverse differ”, I think it’s called?) I’m not sure if you could easily also pull out the near misses? (If you can help me out on how to easily find matching or near matching sentences across documents via a comment or link, it’d be appreciated…:-)

The more general approach is to just measure document similarity – TF-IDF (Term Frequency – Inverse Document Frequency) and cosine similarity are key phrases here. I guess this approach could also be applied to sentences to find common ones across documents, (eg SO: Similarity between two text documents), though I guess it would require comparing quite a large number of sentences (for ~N sentences in each doc, it’d require N^2 comparisons)? I suppose you could optimise by ignoring comparisons between sentences of radically different lengths? Again, presumably there are tools that do this already?

Unlike simply counting common words that aren’t stop words in a document to find the most popular words in a doc, TF-IDF moderates the simple count (the term frequency) with the inverse document frequency. If a word is popular in every document, the term frequency is large and the document frequency is large, so the inverse document frequency (one divided by the document frequency) is small – which in turn gives a reduced TF-IDF value. If a term is popular in one document but not any other, the document frequency is small and so the relative document frequency is large, giving a large TF-IDF for the term in the rare document in which it appears. TF-IDF helps you spot words that are rare across documents or uncommonly frequent within documents.

Topic models: I thought I’d played with these quite a bit before, but if I did the doodles didn’t make it as far as the blog… The idea behind topic modeling is generate a set of key terms – topics – that provide an indication of the topic of a particular document. (It’s a bit more sophisticated than using a count of common words that aren’t stopwords to characterise a document, which is the approach that tends to be used when generating wordclouds…) There are some pointers in the comments to A Quick View Over a MASHe Google Spreadsheet Twitter Archive of UKGC12 Tweets about topic modeling in R using the R topicmodels package; this ROpenSci post on Topic Modeling in R has code for a nice interactive topic explorer; this notebook on Topic Modeling 101 looks like a handy intro to topic modeling using the gensim Python package.

Automatic summarisation/text summary generation: again, I thought I dabbled with this but there’s no sign of it on this blog:-( There are several tools and recipes out there that will generate text summaries of long documents, but I guess they could be hit and miss and I’d need to play with a few of them to see how easy they are to use and how well they seem to work/how useful they appear to be. The python sumy package looks quite interesting in this respect (example usage) and is probably where I’d start. A simple description of a basic text summariser can be found here: Text summarization with NLTK.

So – what have I missed?

PS In passing, see this JISC review from 2012 on the Value and Benefits of Text Mining.