PyCon UK 2016 write-up

Last week I had a long weekend at PyCon UK 2016 in Cardiff, and it’s been a fantastic experience! Great talks, great friends/colleagues and lots of ideas.

On Monday 19th, on the last day of the conference, my friend Miguel and I have run a tutorial/workshop on Natural Language Processing in Python (the GitHub repo contains the Jupyter notebooks we used as well as some slides for an introduction).

Our NLP tutorial

Since I’ve already mentioned it, I’ll start from the end :)

The tutorial was tailored for NLP beginners and, as I mentioned explicitly at the very beginning, I wasn’t there to impress the experts. Rather, the whole point was to get the attendees a bit curious about Natural Language Processing, and to show them what you can do with a few lines of Python.

Overall, I think we’ve been quite lucky as we had the perfect audience: the right number of people (around 20+) with a bit of Python knowledge but not much NLP knowledge.

We only had some minor hiccups with the installation process, which is something we’re going to work on to make it smoother and more beginner-friendly. In particular the things I’d like to improve are:

  • add some testing / pre-flight checks, e.g. “how do I know that the environment is set up correctly?” (Miguel has already added this)
  • support for Windows: I’m quite useless with trouble-shooting Windows issues, but a couple of attendees had some troubles with the installation process not going too smoothly; maybe some virtual machine setup will be helpful

I also think having the material available in advance, so the attendees can start setting up the environment is very helpful. Most of them were quite engaged and I received a couple of “bug reports” on-the-fly, even a pull request that improved the installation process (thanks!)

Last but not least, I was also happy to give out a copy of my book (Mastering Social Media Mining with Python) that I had with me (the raffle was implemented on the spot through random.choice(), and the book went to Paivi from Django Girls).

I’ll give a shorter version of this tutorial at PyCon Ireland later this year, so in case you’ll be around, I’ll see you there :)

Unfortunately, the tutorials were not recorded so there is no video on-line, but the slides are in the GitHub repo so please dig in and send feedback if you have any.

The Open Day

Thursday 15th was “day zero” of the conference, hosted at Cardiff University. The ticket was free, although there was limited capacity. The day was aimed at introducing the new audience to Python and PyCon. We haven’t seen much Python code on that day, as the talks were mainly for newcomers, yet we had a lot of food for thoughs. This is a great way to introduce more people to Python and to show them how the community is friendly and happy to get more beginners on board.

Teachers, Kids and Education

One of the main themes of the conference was Education. Friday 16th, the first day of the main event, was labelled “Teachers Day”, while Saturday 17th was “Kids Day”. The effort to make CS education more accessible for kids was very clear, and some of the initiatives were really spot-on. In particular, some of the kids have been able to hack some small project together in a very short time, and they delivered a “show and tell” session at the end of the second day. I think their creativity and the fact that they were standing in front of a crowd of 500+ developers to show what they have been working on during their day have been very impressive.

Community in the Broader Sense

Another aspect that became quite clear is the strength of the Python Community. Some representatives of PyCon Poland, PyCon Switzerland and Django Europe were introducing their upcoming events. Some attendees with less economic capabilities were given the opportunity to attend, through some form of financial support (including e.g. students from India).

Representatives from PyCon Namibia and PyCon Zimbabwe were also attending and they discussed some of the challenges they are facing while building a local community in their countries.

In particular, the work Jessica from PyNAM is carrying out with young learners is extremely inspiring and deserves more visibility (link to the video of her talk).

Accessibility for Everybody

One of the features that I’ve never experienced in a conference so far was the speech-to-text transcription. During the talks, the speech-to-text team have been very busy writing down what the speakers were saying in real-time. While this is sometimes considered an accessibility feature which might benefit only deaf users, it turned out live captions are extremely beneficial for everybody. Firstly, not all the non-deaf attendees have perfect hearing. Secondly, not everybody is an English native speaker (both speakers and audience), so a word might be missed, or an accent might cause some confusion. Lastly, not every attendee is paying full attention to every talk for the whole talk: sometimes towards the end of the day, you just switch off for a moment and the live captions allow you to catch up.

Providing some accessibility feature turned out to be beneficial for everybody.

Shout out to the Organisers

Organising such a big event (500+ attendees) is not an easy task, so all the people who have worked hard to make this conference happen deserve a big round of applause. Not naming names here, but if you’ve been involved, thanks!

Being Interviewed about NLP

This was a bit random, in a very pleasant way. On Saturday, Miguel, Lev from RaRe Technologies and I spent some time with Kate Jarmul, who by the way just introduced her book on data wrangling, and also delivered a tutorial on the topic. The topic of the conversation was on our views, in the broader sense, about NLP / Text Analytics, how we got into this field, how we see this field evolving and so on. Apparently, this was an interview with some experts of the field, for a piece she’s writing for the O’Reilly blog (I should put an amazed emoticon here).

Using Python for …

The breadth of the topics discussed during the conference was really amazing. I think this kind of events are a great way to see what people are working on and how the tools we use every day are used by other people.

I’m not going to name any talk in particular, because there are too many good talks that deserve to be mentioned.

In terms of topics, some fields that are well covered by Python are:

  • Data Science (and related topics like data cleaning, NLP and machine learning)
  • Web development (with Django and so many interesting libraries)
  • electronics and robotics (with Raspberry Pi, micro:bit, MicroPython etc)
  • you name it :)

I’m probably not saying anything new here, but it was nice to see it in first person and step outside my data-sciency comfort zone.


Thanks to everybody who contributed to this event, and see you in Cardiff for PyCon UK 2017!

Mastering Social Media Mining with Python


Great news, my book on data mining for social media is finally out!

The title is Mastering Social Media Mining with Python. I’ve been working with Packt Publishing over the past few months, and in July the book has been finalised and released.


As part of Packt’s Mastering series, the book assumes the readers already have some basic understanding of Python (e.g. for loops and classes), but more advanced concepts are discussed with examples. No particular experience with Social Media APIs and Data Mining is required. With 300+ pages, by the end of the book, the readers should be able to build their own data mining projects using data from social media and Python tools.

A bird’s eye view on the content:

  1. Social Media, Social Data and Python
    • Introduction on Social Media and Social Data: challenges and opportunities
    • Introduction on Python tools for Data Science
    • Overview on the use of public APIs to interact with social media platforms
  2. #MiningTwitter: Hashtags, Topics and Time Series
    • Interacting with the Twitter API in Python
    • Twitter data: the anatomy of a tweet
    • Entity analysis, text analysis, time series analysis on tweets
  3. Users, Followers, and Communities on Twitter
    • Analysing who follows whom
    • Mining your followers
    • Mining communities
    • Visualising tweets on a map
  4. Posts, Pages and User Interactions on Facebook
    • Interacting the Facebook Graph API in Python
    • Mining you posts
    • Mining Facebook Pages
  5. Topic analysis on Google Plus
    • Interacting with the Google Plus API in Python
    • Finding people and pages on G+
    • Analysis of notes and activities on G+
  6. Questions and Answers on Stack Exchange
    • Interacting with the StackOverflow API in Python
    • Text classification for question tags
  7. Blogs, RSS, Wikipedia, and Natural Language Processing
    • Blogs and web pages as social data Web scraping with Python
    • Basics of text analytics on blog posts
    • Information extraction from text
  8. Mining All the Data!
    • Interacting with many other APIs and types of objects
    • Examples of interaction with YouTube, Yelp and GitHub
  9. Linked Data and the Semantic Web
    • The Web as Social Media
    • Mining relations from DBpedia
    • Mining geo coordinates

The detailed table of contents is shown on the Packt Pub’s page. Chapter 2 is also offered as free sample.

Please have a look at the companion code for the book on my GitHub, so you can have an idea of the applications discussed in the book.

A Brief Introduction to Text Summarisation

In this article, I’ll discuss some aspects of text summarisation, the process of analysing a text document, or a set of documents, in order to produce a summary of its content. The overall purpose is to reduce the amount of information that a user has to digest in order to understand whether reading the whole document is relevant for its information need.

This article is a bird’s-eye view on the topic, to understand the different implications of the problem, rather than a detailed discussion on a specific implementation. The latter will be the subject of future articles.

Summarisation is one of the important tasks in text analytics and it’s an active area of (academic) research which involves mainly the Natural Language Processing and Information Retrieval communities.

Information Overload and the Need for a Good Summary

The core of the matter is the information overload we are experiencing on a daily basis. To put it simply, there is just too much information to digest, and not enough time to do it. The purpose of summarisation is to minimise the amount of information you have to go through, before you can grasp the overall concepts described in the document.

Summarisation can happen in different forms, but the key idea is to present the user with something short, yet informative.

To name just one example, let’s say you want to buy some product, and you’d like to get some opinions about such product. None of your friends owns the product, so you have a look at the on-line reviews: thousands and thousands of sentences to read. Are you going to read all of them? Do you read just some of them and hope to get the best insights?

This is how Google Shopping provides the user with a possible solution:

Example of Google Shopping Summary
Example of Summary from Google Shopping

In this image, the reviews about a popular gaming console are condensed, providing a distribution of ratings and a breakdown of different aspects about the product (e.g. picture/video or battery). The user can then decide to read further, by clicking on a specific aspect, or on a specific rating. Other popular on-line services offer similar

Maybe this is not a big issue when the value of the product is just a few pounds/dollas/euros, but the same problem will arise any time there is just too much to read, and not enough time.

Application Scenarios

As mentioned in the previous paragraph, every scenario where there is a lot of text to read can be a good application scenario for text summarisation. The following list is paraphrased from a tutorial given at ACL 2001 by Maybury and Many:

  • News summaries: what happened while I was away?
  • Physicians’ aids: summarise and compare the recommended treatments for this patient
  • Meeting summarisation: what happended at the meeting I missed?
  • Search engine result pages: snippets of the retrieved documents compared to the query
  • Intelligence: create 500-word file of a suspect
  • Small screens: create a screen-sized summary of a book/email
  • Aids for visually impaired: compact the text and read it out for a blind person

More examples:

  • Sentiment Analysis: give me pros and cons of a product
  • Social media: what are the trending topics today?

Text summarisation is not the only way to tackle the information overload in some of these scenarios, but it can play an important role and it can be used as a component of a more complex system that involves e.g. recommendations and search.

Properties of a Summary

Before we can build a summarisation system, we need to understand how the summary is going to be consumed.

There are many different ways to characterise a summary, here we summarise some of them.

Abstract vs. extract: do we rephrase the content or do we extract some of it? The first involves natural language generation, the latter involves e.g. phrase/sentence ranking.

Single vs Multi source: multiple sources can introduce discrepancies, confirming or contradicting some information. Reviews stating opposite opinions and experiences can be legitimate. News releases that contradict each others are problematic. How do we deal with duplicate content? How do we promote novelty?

Generic vs User-oriented: a generic summary is static, created once for all the users. A user-oriented summary is dynamic, tailored to a particular user profile or user session.

Summary Function: do we want to cover all the key points of the source, or just act as a preview? Do we provide an additional critical view on the source? Think about a movie, and compare its plot, its trailer and a review about it: they are all summaries of the movie, but they have different functions.

Summary Length: a summary should be… short. How short? Do we have a target length (number of words/sentences) or a compression rate (e.g. 5% of the source)?

Linguistic qualities: is the summary coherent? Is it fluent? Is it self-contained?

These are just some of the aspects to consider when building a summarisation application. The key question probably is: how is it going to help the user?

Evaluation: How Good is a Summary?

Evaluating a summary is a challenge in itself. The previous paragraph has opened the discussion for a variety of summarisation approaches, so in order to decide how good a summary is, we really need to put some more context.

In principle, there two orthogonal ways to evaluate summarisation: user-vs-system-based and intrinsic-vs-extrinsic. Let’s briefly discuss them.

User-based intrinsic: users are asked to judge the quality of the summary per se. A typical question could be as simple as “how did you like the summary?“, or something more complex regarding the coherence of the summary, or whether it was helpful to understand the full text.

User-based extrinsic: users are asked to complete a particular task. The quality of the summary is measured against how well the user performs on the task. Here, “how well” can involve, for example, accuracy or speed: does the summary improve the user’s performance?

System-based intrinsic: gold standard summaries are produced by human judges, and the system-generated summaries are compared against them. Some evaluation metrics are involved for the automatic generation of a score that allows summarisation systems to be compared. A common example is the ROUGE framework, based on n-gram overlaps.

System-based extrinsic: the system performs some other tasks (e.g. text classification), using the system-generated summary. The system performances are evaluated for these other tasks, with and without the use of the summarisation component.

In general, involving users is a longer and more expensive process, but can provide interesting insights in terms of summary quality. System-based summarisation with e.g. ROUGE can be useful for some initial comparison, when many potential system candidates are available and employing users to judge all of them could be simply not feasible.

Evaluating a summariser against a particular task (extrinsic evaluation) often helps to answer the initial question, how is the summary going to help the user?


To summarise :) there are a few aspects to consider before building a summarisation system.

This article has provided an overall introduction to the field, to highlight some of the key issues to think about.

Some follow-up article will provide more concrete examples with existing tools and actual implementation, to showcase the real use of text summarisation.

Mining Twitter Data with Python (Part 6 – Sentiment Analysis Basics)

Sentiment Analysis is one of the interesting applications of text analytics. Although the term is often associated with sentiment classification of documents, broadly speaking it refers to the use of text analytics approaches applied to the set of problems related to identifying and extracting subjective material in text sources.

This article continues the series on mining Twitter data with Python, describing a simple approach for Sentiment Analysis and applying it to the rubgy data set (see Part 4).

Tutorial Table of Contents:

A Simple Approach for Sentiment Analysis

The technique we’re discussing in this post has been elaborated from the traditional approach proposed by Peter Turney in his paper Thumbs Up or Thumbs Down? Semantic Orientation Applied to Unsupervised Classification of Reviews. A lot of work has been done in Sentiment Analysis since then, but the approach has still an interesting educational value. In particular, it is intuitive, simple to understand and to test, and most of all unsupervised, so it doesn’t require any labelled data for training.

Firstly, we define the Semantic Orientation (SO) of a word as the difference between its associations with positive and negative words. In practice, we want to calculate “how close” a word is with terms like good and bad. The chosen measure of “closeness” is Pointwise Mutual Information (PMI), calculated as follows (t1 and t2 are terms):

\mbox{PMI}(t_1, t_2) = \log\Bigl(\frac{P(t_1 \wedge t_2)}{P(t_1) \cdot P(t_2)}\Bigr)

In Turney’s paper, the SO of a word was calculated against excellent and poor, but of course we can extend the vocabulary of positive and negative terms. Using V^{+} and a vocabulary of positive terms and V^{-} for the negative ones, the Semantic Orientation of a term t is hence defined as:

\mbox{SO}(t) = \sum_{t' \in V^{+}}\mbox{PMI}(t, t') - \sum_{t' \in V^{-}}\mbox{PMI}(t, t')

We can build our own list of positive and negative terms, or we can use one of the many resources available on-line, for example the opinion lexicon by Bing Liu.

Computing Term Probabilities

In order to compute P(t) (the probability of observing the term t) and P(t_1 \wedge t_2) (the probability of observing the terms t1 and t2 occurring together) we can re-use some previous code to calculate term frequencies and term co-occurrences. Given the set of documents (tweets) D, we define the Document Frequency (DF) of a term as the number of documents where the term occurs. The same definition can be applied to co-occurrent terms. Hence, we can define our probabilities as:

P(t) = \frac{\mbox{DF}(t)}{|D|}\\  P(t_1 \wedge t_2) = \frac{\mbox{DF}(t_1 \wedge t_2)}{|D|}

In the previous articles, the document frequency for single terms was stored in the dictionaries count_single and count_stop_single (the latter doesn’t store stop-words), while the document frequency for the co-occurrencies was stored in the co-occurrence matrix com

This is how we can compute the probabilities:

# n_docs is the total n. of tweets
p_t = {}
p_t_com = defaultdict(lambda : defaultdict(int))

for term, n in count_stop_single.items():
    p_t[term] = n / n_docs
    for t2 in com[term]:
        p_t_com[term][t2] = com[term][t2] / n_docs

Computing the Semantic Orientation

Given two vocabularies for positive and negative terms:

positive_vocab = [
    'good', 'nice', 'great', 'awesome', 'outstanding',
    'fantastic', 'terrific', ':)', ':-)', 'like', 'love',
    # shall we also include game-specific terms?
    # 'triumph', 'triumphal', 'triumphant', 'victory', etc.
negative_vocab = [
    'bad', 'terrible', 'crap', 'useless', 'hate', ':(', ':-(',
    # 'defeat', etc.

We can compute the PMI of each pair of terms, and then compute the
Semantic Orientation as described above:

pmi = defaultdict(lambda : defaultdict(int))
for t1 in p_t:
    for t2 in com[t1]:
        denom = p_t[t1] * p_t[t2]
        pmi[t1][t2] = math.log2(p_t_com[t1][t2] / denom)

semantic_orientation = {}
for term, n in p_t.items():
    positive_assoc = sum(pmi[term][tx] for tx in positive_vocab)
    negative_assoc = sum(pmi[term][tx] for tx in negative_vocab)
    semantic_orientation[term] = positive_assoc - negative_assoc

The Semantic Orientation of a term will have a positive (negative) value if the term is often associated with terms in the positive (negative) vocabulary. The value will be zero for neutral terms, e.g. the PMI’s for positive and negative balance out, or more likely a term is never observed together with other terms in the positive/negative vocabularies.

We can print out the semantic orientation for some terms:

semantic_sorted = sorted(semantic_orientation.items(), 
top_pos = semantic_sorted[:10]
top_neg = semantic_sorted[-10:]

print("ITA v WAL: %f" % semantic_orientation['#itavwal'])
print("SCO v IRE: %f" % semantic_orientation['#scovire'])
print("ENG v FRA: %f" % semantic_orientation['#engvfra'])
print("#ITA: %f" % semantic_orientation['#ita'])
print("#FRA: %f" % semantic_orientation['#fra'])
print("#SCO: %f" % semantic_orientation['#sco'])
print("#ENG: %f" % semantic_orientation['#eng'])
print("#WAL: %f" % semantic_orientation['#wal'])
print("#IRE: %f" % semantic_orientation['#ire'])

Different vocabularies will produce different scores. Using the opinion lexicon from Bing Liu, this is what we can observed on the Rugby data-set:

# the top positive terms
[('fantastic', 91.39950482011552), ('@dai_bach', 90.48767241244532), ('hoping', 80.50247748725415), ('#it', 71.28333427277785), ('days', 67.4394844955977), ('@nigelrefowens', 64.86112716005566), ('afternoon', 64.05064208341855), ('breathtaking', 62.86591435212975), ('#wal', 60.07283361352875), ('annual', 58.95378954406133)]
# the top negative terms
[('#england', -74.83306534609066), ('6', -76.0687215594536), ('#itavwal', -78.4558633116863), ('@rbs_6_nations', -80.89363516601993), ("can't", -81.75379628180468), ('like', -83.9319149443813), ('10', -85.93073078165587), ('italy', -86.94465165178258), ('#engvfra', -113.26188957010228), ('ball', -161.82146824640125)]
# Matches
ITA v WAL: -78.455863
SCO v IRE: -73.487661
ENG v FRA: -113.261890
# Individual team
#ITA: 53.033824
#FRA: 14.099372
#SCO: 4.426723
#ENG: -0.462845
#WAL: 60.072834
#IRE: 19.231722

Some Limitations

The PMI-based approach has been introduced as simple and intuitive, but of course it has some limitations. The semantic scores are calculated on terms, meaning that there is no notion of “entity” or “concept” or “event”. For example, it would be nice to aggregate and normalise all the references to the team names, e.g. #ita, Italy and Italia should all contribute to the semantic orientation of the same entity. Moreover, do the opinions on the individual teams also contribute to the overall opinion on a match?

Some aspects of natural language are also not captured by this approach, more notably modifiers and negation: how do we deal with phrases like not bad (this is the opposite of just bad) or very good (this is stronger than just good)?


This article has continued the tutorial on mining Twitter data with Python introducing a simple approach for Sentiment Analysis, based on the computation of a semantic orientation score which tells us whether a term is more closely related to a positive or negative vocabulary. The intuition behind this approach is fairly simple, and it can be implemented using Pointwise Mutual Information as a measure of association. The approach has of course some limitations, but it’s a good starting point to get familiar with Sentiment Analysis.


Mining Twitter Data with Python: Part 5 – Data Visualisation Basics

A picture is worth a thousand tweets: more often than not, designing a good visual representation of our data, can help us make sense of them and highlight interesting insights. After collecting and analysing Twitter data, the tutorial continues with some notions on data visualisation with Python.

Tutorial Table of Contents:

From Python to Javascript with Vincent

While there are some options to create plots in Python using libraries like matplotlib or ggplot, one of the coolest libraries for data visualisation is probably D3.js which is, as the name suggests, based on Javascript. D3 plays well with web standards like CSS and SVG, and allows to create some wonderful interactive visualisations.

Vincent bridges the gap between a Python back-end and a front-end that supports D3.js visualisation, allowing us to benefit from both sides. The tagline of Vincent is in fact “The data capabilities of Python. The visualization capabilities of JavaScript”. Vincent, a Python library, takes our data in Python format and translates them into Vega, a JSON-based visualisation grammar that will be used on top of D3. It sounds quite complicated, but it’s fairly simple and pythonic. You don’t have to write a line in Javascript/D3 if you don’t want to.

Firstly, let’s install Vincent:

sudo pip install vincent

Secondly, let’s create our first plot. Using the list of most frequent terms (without hashtags) from our rugby data set, we want to plot their frequencies:

import vincent

word_freq = count_terms_only.most_common(20)
labels, freq = zip(*word_freq)
data = {'data': freq, 'x': labels}
bar = vincent.Bar(data, iter_idx='x')

At this point, the file term_freq.json will contain a description of the plot that can be handed over to D3.js and Vega. A simple template (taken from Vincent resources) to visualise the plot:

    <title>Vega Scaffold</title>
    <script src="" charset="utf-8"></script>
    <script src=""></script>
    <script src="" charset="utf-8"></script>
    <script src=""></script>
    <div id="vis"></div>
<script type="text/javascript">
// parse a spec and create a visualization view
function parse(spec) {
  vg.parse.spec(spec, function(chart) { chart({el:"#vis"}).update(); });

Save the above HTML page as chart.html and run the simple Python web server:

python -m http.server 8888 # Python 3
python -m SimpleHTTPServer 8888 # Python 2

Now you can open your browser at http://localhost:8888/chart.html and observe the result:

Term Frequencies

Notice: you could save the HTML template directly from Python with:

bar.to_json('term_freq.json', html_out=True, html_path='chart.html')

but, at least in Python 3, the output is not a well formed HTML and you’d need to manually strip some characters.

With this procedure, we can plot many different types of charts with Vincent. Let’s take a moment to browse the docs and see its capabilities.

Time Series Visualisation

Another interesting aspect of analysing data from Twitter is the possibility to observe the distribution of tweets over time. In other words, if we organise the frequencies into temporal buckets, we could observe how Twitter users react to real-time events.

One of my favourite tools for data analysis with Python is Pandas, which also has a fairly decent support for time series. As an example, let’s track the hashtag #ITAvWAL to observe what happened during the first match.

Firstly, if we haven’t done it yet, we need to install Pandas:

sudo pip install pandas

In the main loop which reads all the tweets, we simply track the occurrences of the hashtag, i.e. we can refactor the code from the previous episodes into something similar to:

import pandas
import json

dates_ITAvWAL = []
# f is the file pointer to the JSON data set
for line in f:
    tweet = json.loads(line)
    # let's focus on hashtags only at the moment
    terms_hash = [term for term in preprocess(tweet['text']) if term.startswith('#')]
    # track when the hashtag is mentioned
    if '#itavwal' in terms_hash:

# a list of "1" to count the hashtags
ones = [1]*len(dates_ITAvWAL)
# the index of the series
idx = pandas.DatetimeIndex(dates_ITAvWAL)
# the actual series (at series of 1s for the moment)
ITAvWAL = pandas.Series(ones, index=idx)

# Resampling / bucketing
per_minute = ITAvWAL.resample('1Min', how='sum').fillna(0)

The last line is what allows us to track the frequencies over time. The series is re-sampled with intervals of 1 minute. This means all the tweets falling within a particular minute will be aggregated, more precisely they will be summed up, given how='sum'. The time index will not keep track of the seconds anymore. If there is no tweet in a particular minute, the fillna() function will fill the blanks with zeros.

To put the time series in a plot with Vincent:

time_chart = vincent.Line(ITAvWAL)
time_chart.axis_titles(x='Time', y='Freq')

Once you embed the time_chart.json file into the HTML template discussed above, you’ll see this output:

Time Series

The interesting moments of the match are observable from the spikes in the series. The first spike just before 1pm corresponds to the first Italian try. All the other spikes between 1:30 and 2:30pm correspond to Welsh tries and show the Welsh dominance during the second half. The match was over by 2:30, so after that Twitter went quiet.

Rather than just observing one sequence at a time, we could compare different series to observe how the matches has evolved. So let’s refactor the code for the time series, keeping track of the three different hashtags #ITAvWAL, #SCOvIRE and #ENGvFRA into the corresponding pandas.Series.

# all the data together
match_data = dict(ITAvWAL=per_minute_i, SCOvIRE=per_minute_s, ENGvFRA=per_minute_e)
# we need a DataFrame, to accommodate multiple series
all_matches = pandas.DataFrame(data=match_data,
# Resampling as above
all_matches = all_matches.resample('1Min', how='sum').fillna(0)

# and now the plotting
time_chart = vincent.Line(all_matches[['ITAvWAL', 'SCOvIRE', 'ENGvFRA']])
time_chart.axis_titles(x='Time', y='Freq')

And the output:


We can immediately observe when the different matches took place (approx 12:30-2:30, 2:30-4:30 and 5-7) and we can see how the last match had the all the attentions, especially in the end when the winner was revealed.


Data visualisation is an important discipline in the bigger context of data analysis. By supporting visual representations of our data, we can provide interesting insights. We have discussed a relatively simple option to support data visualisation with Python using Vincent. In particular, we have seen how we can easily bridge the gap between Python and a language like Javascript that offers a great tool like D3.js, one of the most important libraries for interactive visualisation. Overall, we have just scratched the surface of data visualisation, but as a starting point this should be enough to get some nice ideas going. The nature of Twitter as a medium has also encouraged a quick look into the topic of time series analysis, allowing us to mention pandas as a great Python tool.

If this article has given you some ideas for data visualisation, please leave a comment below or get in touch.


Tutorial Table of Contents:

Mining Twitter Data with Python (Part 4: Rugby and Term Co-occurrences)

Last Saturday was the closing day of the Six Nations Championship, an annual international rugby competition. Before turning on the TV to watch Italy being trashed by Wales, I decided to use this event to collect some data from Twitter and perform some exploratory text analysis on something more interesting than the small list of my tweets.

This article continues the tutorial on Twitter Data Mining, re-using what we discussed in the previous articles with some more realistic data. It also expands the analysis by introducing the concept of term co-occurrence.

Tutorial Table of Contents:

The Application Domain

As the name suggests, six teams are involved in the competition: England, Ireland, Wales, Scotland, France and Italy. This means that we can expect the event to be tweeted in multiple languages (English, French, Italian, Welsh, Gaelic, possibly other languages as well), with English being the major language. Assuming the team names will be mentioned frequently, we could decide to look also for their nicknames, e.g. Les Bleus for France or Azzurri for Italy. During the last day of the competition, three matches are played sequentially. Three teams in particular had a shot for the title: England, Ireland and Wales. At the end, Ireland won the competition but everything was open until the very last minute.

Setting Up

I used the streaming API to download all the tweets containing the string #rbs6nations during the day. Obviously not all the tweets about the event contained the hashtag, but this is a good baseline. The time frame for the download was from around 12:15PM to 7:15PM GMT, that is from about 15 minutes before the first match, to about 15 minutes after the last match was over. At the end, more than 18,000 tweets have been downloaded in JSON format, making for about 75Mb of data. This should be small enough to quickly do some processing in memory, and at the same time big enough to observe something possibly interesting.

The textual content of the tweets has been pre-processed with tokenisation and lowercasing using the preprocess() function introduced in Part 2 of the tutorial.

Interesting terms and hashtags

Following what we discussed in Part 3 (Term Frequencies), we want to observe the most common terms and hashtags used during day. If you have followed the discussion about creating different lists of tokens in order to capture terms without hashtags, hashtags only, removing stop-words, etc. you can play around with the different lists.

This is the unsurprising list of top 10 most frequent terms (terms_only in Part 3) in the data set.

[('ireland', 3163), ('england', 2584), ('wales', 2271), ('…', 2068), ('day', 1479), ('france', 1380), ('win', 1338), ('rugby', 1253), ('points', 1221), ('title', 1180)]

The first three terms correspond to the teams who had a go for the title. The frequencies also respect the order in the final table. The fourth term is instead a punctuation mark that we missed and didn’t include in the list of stop-words. This is because string.punctuation only contains ASCII symbols, while here we’re dealing with a unicode character. If we dig into the data, there will be more examples like this, but for the moment we don’t worry about it.

After adding the suspension-points symbol to the list of stop-words, we have a new entry at the end of the list:

[('ireland', 3163), ('england', 2584), ('wales', 2271), ('day', 1479), ('france', 1380), ('win', 1338), ('rugby', 1253), ('points', 1221), ('title', 1180), ('🍀', 1154)]

Interestingly, a new token we didn’t account for, an Emoji symbol (in this case, the Irish Shamrock).

If we have a look at the most common hashtags, we need to consider that #rbs6nations will be by far the most common token (that’s our search term for downloading the tweets), so we can exclude it from the list. This leave us with:

[('#engvfra', 1701), ('#itavwal', 927), ('#rugby', 880), ('#scovire', 692), ('#ireland', 686), ('#angfra', 554), ('#xvdefrance', 508), ('#crunch', 500), ('#wales', 446), ('#england', 406)]

We can observe that the most common hashtags, a part from #rugby, are related to the individual matches. In particular England v France has received the highest number of mentions, probably being the last match of the day with a dramatic finale. Something interesting to notice is that a fair amount of tweets also contained terms in French: the count for #angfra should in fact be added to #engvfra. Those unfamiliar with rugby probably wouldn’t recognise that also #crunch should be included with #EngvFra match, as Le Crunch is the traditional name for this event. So by far, the last match has received a lot of attention.

Term co-occurrences

Sometimes we are interested in the terms that occur together. This is mainly because the context gives us a better insight about the meaning of a term, supporting applications such as word disambiguation or semantic similarity. We discussed the option of using bigrams in the previous article, but we want to extend the context of a term to the whole tweet.

We can refactor the code from the previous article in order to capture the co-occurrences. We build a co-occurrence matrix com such that com[x][y] contains the number of times the term x has been seen in the same tweet as the term y:

from collections import defaultdict
# remember to include the other import from the previous post

com = defaultdict(lambda : defaultdict(int))

# f is the file pointer to the JSON data set
for line in f: 
    tweet = json.loads(line)
    terms_only = [term for term in preprocess(tweet['text']) 
                  if term not in stop 
                  and not term.startswith(('#', '@'))]

    # Build co-occurrence matrix
    for i in range(len(terms_only)-1):            
        for j in range(i+1, len(terms_only)):
            w1, w2 = sorted([terms_only[i], terms_only[j]])                
            if w1 != w2:
                com[w1][w2] += 1

While building the co-occurrence matrix, we don’t want to count the same term pair twice, e.g. com[A][B] == com[B][A], so the inner for loop starts from i+1 in order to build a triangular matrix, while sorted will preserve the alphabetical order of the terms.

For each term, we then extract the 5 most frequent co-occurrent terms, creating a list of tuples in the form ((term1, term2), count):

com_max = []
# For each term, look for the most common co-occurrent terms
for t1 in com:
    t1_max_terms = sorted(com[t1].items(), key=operator.itemgetter(1), reverse=True)[:5]
    for t2, t2_count in t1_max_terms:
        com_max.append(((t1, t2), t2_count))
# Get the most frequent co-occurrences
terms_max = sorted(com_max, key=operator.itemgetter(1), reverse=True)

The results:

[(('6', 'nations'), 845), (('champions', 'ireland'), 760), (('nations', 'rbs'), 742), (('day', 'ireland'), 731), (('ireland', 'wales'), 674)]

This implementation is pretty straightforward, but depending on the data set and on the use of the matrix, one might want to look into tools like scipy.sparse for building a sparse matrix.

We could also look for a specific term and extract its most frequent co-occurrences. We simply need to modify the main loop including an extra counter, for example:

search_word = sys.argv[1] # pass a term as a command-line argument
count_search = Counter()
for line in f:
    tweet = json.loads(line)
    terms_only = [term for term in preprocess(tweet['text']) 
                  if term not in stop 
                  and not term.startswith(('#', '@'))]
    if search_word in terms_only:
print("Co-occurrence for %s:" % search_word)

The outcome for “ireland”:

[('champions', 756), ('day', 727), ('nations', 659), ('wales', 654), ('2015', 638), ('6', 613), ('rbs', 585), ('', 559), ('🍀', 526), ('10', 522), ('win', 377), ('england', 377), ('twickenham', 361), ('40', 360), ('points', 356), ('sco', 355), ('ire', 355), ('title', 346), ('scotland', 301), ('turn', 295)]

The outcome for “rugby”:

[('day', 476), ('game', 160), ('ireland', 143), ('england', 132), ('great', 105), ('today', 104), ('best', 97), ('well', 90), ('ever', 89), ('incredible', 87), ('amazing', 84), ('done', 82), ('amp', 71), ('games', 66), ('points', 64), ('monumental', 58), ('strap', 56), ('world', 55), ('team', 55), ('', 53)]

Overall, quite interesting.


This article has discussed a toy example of Text Mining on Twitter, using some realistic data taken during a sport event. Using what we have learnt in the previous episodes, we have downloaded some data using the streaming API, pre-processed the data in JSON format and extracted some interesting terms and hashtags from the tweets. The article has also introduced the concept of term co-occurrence, shown how to build a co-occurrence matrix and discussed how to use it to find some interesting insight.


Mining Twitter Data with Python (Part 3: Term Frequencies)

This is the third part in a series of articles about data mining on Twitter. After collecting data and pre-processing some text, we are ready for some basic analysis. In this article, we’ll discuss the analysis of term frequencies to extract meaningful terms from our tweets.

Tutorial Table of Contents:

Counting Terms

Assuming we have collected a list of tweets (see Part 1 of the tutorial), the first exploratory analysis that we can perform is a simple word count. In this way, we can observe what are the terms most commonly used in the data set. In this example, I’ll use the set of my tweets, so the most frequent words should correspond to the topics I discuss (not necessarily, but bear with be for a couple of paragraphs).

We can use a custom tokeniser to split the tweets into a list of terms. The following code uses the preprocess() function described in Part 2 of the tutorial, in order to capture Twitter-specific aspects of the text, such as #hashtags, @-mentions, emoticons and URLs. In order to keep track of the frequencies while we are processing the tweets, we can use collections.Counter() which internally is a dictionary (term: count) with some useful methods like most_common():

import operator 
import json
from collections import Counter

fname = 'mytweets.json'
with open(fname, 'r') as f:
    count_all = Counter()
    for line in f:
        tweet = json.loads(line)
        # Create a list with all the terms
        terms_all = [term for term in preprocess(tweet['text'])]
        # Update the counter
    # Print the first 5 most frequent words

The above code will produce some unimpressive results:

[(':', 44), ('rt', 26), ('to', 26), ('and', 25), ('on', 22)]

As you can see, the most frequent words (or should I say, tokens), are not exactly meaningful.

Removing stop-words

In every language, some words are particularly common. While their use in the language is crucial, they don’t usually convey a particular meaning, especially if taken out of context. This is the case of articles, conjunctions, some adverbs, etc. which are commonly called stop-words. In the example above, we can see three common stop-words – to, and and on. Stop-word removal is one important step that should be considered during the pre-processing stages. One can build a custom list of stop-words, or use available lists (e.g. NLTK provides a simple list for English stop-words).

Given the nature of our data and our tokenisation, we should also be careful with all the punctuation marks and with terms like RT (used for re-tweets) and via (used to mention the original author of an article or a re-tweet), which are not in the default stop-word list.

from nltk.corpus import stopwords
import string

punctuation = list(string.punctuation)
stop = stopwords.words('english') + punctuation + ['rt', 'via']

We can now substitute the variable terms_all in the first example with something like:

terms_stop = [term for term in preprocess(tweet['text']) if term not in stop]

After counting, sorting the terms and printing the top 5, this is the result:

[('python', 11), ('@miguelmalvarez', 9), ('#python', 9), ('data', 8), ('@danielasfregola', 7)]

So apparently I mostly tweet about Python and data, and the users I re-tweet more often are @miguelmalvarez and @danielasfregola, it sounds about right.

More term filters

Besides stop-word removal, we can further customise the list of terms/tokens we are interested in. Here you have some examples that you can embed in the first fragment of code:

# Count terms only once, equivalent to Document Frequency
terms_single = set(terms_all)
# Count hashtags only
terms_hash = [term for term in preprocess(tweet['text']) 
              if term.startswith('#')]
# Count terms only (no hashtags, no mentions)
terms_only = [term for term in preprocess(tweet['text']) 
              if term not in stop and 
              not term.startswith(('#', '@'))] 
              # mind the ((double brackets))
              # startswith() takes a tuple (not a list) if 
              # we pass a list of inputs

After counting and sorting, these are my most commonly used hashtags:

[('#python', 9), ('#scala', 6), ('#nosql', 4), ('#bigdata', 3), ('#nlp', 3)]

and these are my most commonly used terms:

[('python', 11), ('data', 8), ('summarisation', 6), ('twitter', 5), ('nice', 5)]


While the other frequent terms represent a clear topic, more often than not simple term frequencies don’t give us a deep explanation of what the text is about. To put things in context, let’s consider sequences of two terms (a.k.a. bigrams).

from nltk import bigrams 

terms_bigram = bigrams(terms_stop)

The bigrams() function from NLTK will take a list of tokens and produce a list of tuples using adjacent tokens. Notice that we could use terms_all to compute the bigrams, but we would probably end up with a lot of garbage. In case we decide to analyse longer n-grams (sequences of n tokens), it could make sense to keep the stop-words, just in case we want to capture phrases like “to be or not to be”.

So after counting and sorting the bigrams, this is the result:

[(('nice', 'article'), 4), (('extractive', 'summarisation'), 4), (('summarisation', 'sentence'), 3), (('short', 'paper'), 3), (('paper', 'extractive'), 2)]

So apparently I tweet about nice articles (I wouldn’t bother sharing the boring ones) and extractive summarisation (the topic of my PhD dissertation). This also sounds about right.


This article has built on top of the previous ones to discuss some basis for extracting interesting terms from a data set of tweets, by using simple term frequencies, stop-word removal and n-grams. While these approaches are extremely simple to implement, they are quite useful to have a bird’s eye view on the data. We have used some components of NLTK (introduced in a previous article), so we don’t have to re-invent the wheel.


Tutorial Table of Contents: