I am currently in the process of completing my masters dissertation.  It’s a large document which I have spent the last year or so writing and I am pretty happy with how it has all turned out.  I’ll be sharing more about the contents and findings of the thesis in the next little while.

As I was editing the thesis I thought it might be interesting to try and visualize how referencing occurred within the document.  I had to double check all of the references anyhow, so I thought I would try to make the process more interesting by: programmaticly extracting all of the in text references; creating a list of references and where they occur in the document (by chapter); and then trying to visualize the connections between references and chapters throughout the entire document.

After mentioning this to a number of my colleagues, I decided to document the process as it generated some interest.  Also I find it useful to document the process in case I need to do it again later.  Credits to my colleague Andrew Deacon for helping me formulate this process.

We are going to extract all of the references in the document.  Start by copying all of the text from your document into an advanced text editor such as Notepad++. You can then use the find and replace function within Notepad++ to identify the parenthesis which surrounds each reference. We want to get each reference on its own line so we can generate a list. Open the ‘Find’ utility and turn on the ‘Regular expression’ search mode in the bottom right of the window.

The regular expression search will enable you to search and replace with paragraph breaks.  Start by searching for ‘\(‘ to identify the opening parenthesis.  The backslash is required as an escape character because the regular expression search is turned on.  You can now find and count the number of opening parenthesis in your document.  We want to have each opening parenthesis on a new line, so replace the ‘\(’ with ‘\r\n(‘.

Now we can do the same for the trailing parenthesis.  So do another find and replace this time replace ‘\)’ with ‘)\r\n’.  Now each in text reference should be on its own line.

Now search again this time in normal search mode for an ‘(’, and hit the button ‘Find All in Current Document’.   You should be presented with a list of search results with each in text reference that can be easily copied to Excel.

You will still have to sort out the combined references, et al’s, remove the acronyms, and non-reference parenthesis occurrences.  I also stripped the parenthesis from the text.  Keep the line number so you can determine which chapter each reference occurred within.  If you back to Notepad++ you can grab the line numbers where each section begins and end, mine was quite simple as I only had five chapters.

CH1 lines: 1-132
CH2 133-638
CH3 639-881
CH4 882-1394
CH5 1395-1479

Once you have cleansed the data you should have a clean list of references.  You can use a VLOOKUP to bring in the chapter numbers in Excel.  This process was also very useful for verifying my references and ensuring I used the full reference in a multi author (+3) paper when it first occurred in text.

You can then create a pivot table which will show the references per chapter also showing how many times each paper was referenced.

Now you can take columns A and B shown above and drop them into something like NodeXL to visualize the connections between your reference occurrences (you will have to copy down the ‘CH1’ text).   You can use the ‘Totals’ column to make the connecting lines bigger where a reference was used multiple times.   I use the Total field as the Edge Width connecting the reference to the chapter.   You will also want to make the chapter nodes a larger more distinct object, I have used coloured discs. I have also run a grouping algorithm on the dataset which identifies the references which best group together and applies a colour for each node.

The visual shows the references used in my thesis as they occurred within chapters, and were reused within other chapters.  The central literature which I use to support my thesis ends up in the middle of the image (the central red text).

Looking back, this is quite a useful visual of how I constructed my thesis and how certain resources were woven into the various chapters of my research.  Chapter 2 naturally contains the most references as it forms my literature review.  The literature which ultimately becomes useful in my study is used again in Chapter 5 my conclusion; as well as being mentioned in Chapter 1 as I introduce the study.   Two key resources, Engëstrom, 1987 (theoretical framework) and Cohen, Manion & Morrison, 2007 (research design) are used within Chapter 2 and 3, as I explain and then localise their application in the study.

I am toying with the idea of including this in front of the bibliography in my thesis submission.  Think it might be useful or actually annoy an external examiner?

The most central literature is referenced below:

Conole, G., McAndrew, P. & Dimitriadis, Y. (2010). The role of CSCL pedagogical patterns as mediating artefacts for repurposing Open Educational Resources. In: Pozzi, Francesca and Persico, Donatella eds. Techniques for Fostering Collaboration in Online Learning Communities: Theoretical and Practical Perspectives. Hershey, USA: IGI Global.

Cohen, L., Manion, L & Morrison, K.  (2007). Research methods in education.  6th edition. London: Routledge.

Engeström, Y. (1987). Learning by expanding: An activity-theoretical approach to developmental research. Helsinki: Orienta-Konsultit.

Harley, D., Henke, J., Lawrence, S., Miller, I., Perciali, I., Nasatir, D.  (2006). Use and Users of Digital Resources: A Focus on Undergraduate Education in the Humanities and Social Sciences.  Center for Studies in Higher Education (CSHE), University of California, Berkeley.

Hatakka, M. (2009). Build it and They Will Come? – Inhibiting Factors for Reuse of Open Content in Developing Countries.  The Electronic Journal of Information Systems in Developing Countries. EJISDC (2009) 37, 5, 1-16.

This work was inspired by the “Anatomy of the Osama Tweet“, in which the news of Osama Bin Laden’s death spreads just before it was officially announced by U.S. President Obama.  Naturally the original speculative tweet spread virally through Twitter.  My colleague Andrew Deacon and I have undertaken a similar analysis of information flows during an incident which occurred on the University of Cape Town (UCT) yesterday. With the proliferation and adoption of social networks, information has a tendency to spread through networks faster and more virally.  Now because it is happening electronically we can retrospectively run an analysis of how it all plays out.   But it takes the right kind of incident to actually warrant the viral spread of information.  As this article highlights, most of the information posted to popular social networks actually goes nowhere, in terms of amplification or repetition (93% of tweets go nowhere for example).

So when I heard about the crash landing of a helicopter at UCT I knew we would have a viral situation on our hands.  In fact on the way in to the office, just following the incident, I was already hearing students chatting about it on the city campus.  People chat in corridors, sms messages are sent, phones ring, bbm’s buzz, whatsapp’s beep, Facebook statuses are updated, and Tweets are posted to Twitter.  The latter, Twitter, can be said to be the most open and accessible of networks for analysis.  We collect tweets from Twitter on an ongoing basis whenever someone tweets about #UCT and our dataset is slowly growing.  This incident prompted many people to tweet about UCT and the helicopter crash from all over South Africa and the world.  So whereas only a few years ago one might have only heard about this incident on the six o’clock news, people all over the world were already discovering the story by getting the news from social networks.

We harvested 1168 nodes and 1681 edges worth of data from Twitter at about 3pm on August 3rd, 2011.  The tweets had to have included the words “UCT” and “helicopter” to be included.  That means our analysis includes 1168 distinct people who Tweeted about the incident and within this network of people 1681 were connected in some way.  This connection may be a retweet (amplification) or mention (reference).  The lines show these connections between the nodes (people) tweeting about the incident.  Zoom in with your mouse or the + button in the image to explore.

One can see the twits that were most responsible for the spread of information via Twitter in this visual quite clearly.  Mr_capeTown and GarethCliff are shown as very central nodes in the diagram.  As is the media outlets Radio702, 945Kfm, MyNews24 and UctRadio. There are a host of satellite networks around the outside of the diagram as well, which are not connected to the central network.

If you want to explore this social network graph in higher detail download this PDF file.  Within the PDF you can zoom in ultra close and even search to find your Twitter user name.

Most frequently retweeted or mentioned users are shows in the graph below.

The incident was said to have happened around 9:30 am and the tweets started pouring in fast. The image below shows the number of tweets per five minute interval following the incident.

The first photos were posted by folks who captured the scene on their cell phone cameras.  The two most popular images which are still available here and here.  Gradually, news teams arrived and the following video by official media outlet News24 was posted to YouTube.

Only a portion of users tweets had their location data.  There is an option  in Twitter to have your coordinates included with each tweet.  Based on the data we had available we came up with the following view of tweets originating in South Africa.  The lines again represent retweets and mentions between locations.

Circles representing tweets in the bottom left are from Cape Town, then along the Southern Coast to Port Elizabeth, East London and Durban.  The large cluster diagonally up from Cape Town would be Johannesburg/Pretoria with Bloemfontein in between.

Globally the available location data is shown below.

We can infer that tweets went out from parts of Europe, Asia, North America and other African countries.  In many cases these foreign tweeters were in some way connected to a South Africa tweeter indicating a viral global spread.

Discussion

There are a number of limitation to this small scale real-time study.  We have only captured the conversations which were happening within Twitter in the first case.  Furthermore we only capture tweets which include the words “UCT” and “helicopter”, thus missing tweets which use abbreviations such as “heli” etc.  We are also missing the conversations which exist within the other social networks.  I am sure that the conversations which happened in Facebook, a much more widely adopted social network in South Africa, would also be incredibly interesting to analyze.  We have additionally noticed activity streams on the event happening in LinkedIn.

One of the toughest current hurdles in doing such an analysis is actually getting at the data you want.  In this case we used NodeXL to extract the data from Twitter using the Twitter API.  We then build the network graphs in Gephi while using excel to summarize the data.  One of the reasons we were able to get access to this data was that the event happened suddenly and quickly and was not too massive.  Larger events which unfold over time seem to be more difficult to gather data for at this point.

I think its remarkable that this event happened yesterday and we are already able to analyze the social nature of some of the information flows around the event. The next phase of this analysis is to try and replay the flow of tweets based on their timestamp.  We want to examine exactly how the information flowed through the network over time.  I am surprised that the current toolset we have does not allow this.  Gephi does have a ‘Graph Streaming’ tool, but it has to be hooked up to a live feed of information using JSON.   We are looking for help or suggestions as to how to use this dataset to reconstruct the social spread of information.  Leave us a comment if you have any tips!