Glimpses though the clouds: collocates in a new light

This paper demonstrates a web-based, interactive data
visualisation, allowing users to quickly inspect and browse the
collocational relationships present in a corpus. The software is
inspired by tag clouds, fi rst popularised by on-line photograph
sharing website Flickr (www.fl ickr.com). A paper based on
a prototype of this Collocate Cloud visualisation was given
at Digital Resources for the Humanities and Arts 2007. The
software has since matured, offering new ways of navigating
and inspecting the source data. It has also been expanded to
analyse additional corpora, such as the British National Corpus
(http://www.natcorp.ox.ac.uk/), which will be the focus of this
talk.
Tag clouds allow the user to browse, rather than search for
specifi c pieces of information. Flickr encourages its users to
add tags (keywords) to each photograph uploaded. The tags
associated with each individual photograph are aggregated; the
most frequent go on to make the cloud. The cloud consists of
these tags presented in alphabetical order, with their frequency
displayed as variation in colour, or more commonly font size.
Figure 1 is an example of the most popular tags at Flickr: Figure 1. Flickr tag cloud showing 125 of the
most popular photograph keywords
http://www.fl ickr.com/photos/tags/ (accessed 23 November 2007)
The cloud offers two ways to access the information. If the
user is looking for a specifi c term, the alphabetical ordering
of the information allows it to be quickly located if present.
More importantly, as a tool for browsing, frequent tags stand
out visually, giving the user an immediate overview of the
data. Clicking on a tag name will display all photographs which
contain that tag. The cloud-based visualisation has been successfully applied to
language. McMaster University’s TAPoR Tools (http://taporware.
mcmaster.ca/) features a ‘Word Cloud’ module, currently in
beta testing. WMatrix (http://ucrel.lancs.ac.uk/wmatrix/) can
compare two corpora by showing log-likelihood results in
cloud form. In addition to other linguistic metrics, internet
book seller Amazon provides a word cloud, see figure 2. Figure 2. Amazon.com’s ‘Concordance’ displaying the
100 most frequent words in Romeo and Juliet
http://www.amazon.com/Romeo-Juliet-Folger-Shakespeare-
Library/dp/sitb-next/0743477111/ref=sbx_con/104-
4970220-2133519?ie=UTF8&qid=1179135939&sr=1-
1#concordance (accessed 23 November 2007)
In this instance a word frequency list is the data source, showing
the most frequent 100 words. As with the tag cloud, this list
is alphabetically ordered, the font size being proportionate to
its frequency of usage. It has all the benefi ts of a tag cloud; in
this instance clicking on a word will produce a concordance
of that term.
This method of visualisation and interaction offers another
tool for corpus linguists. As developer for an online corpus
project, I have found that the usability and sophistication of
our tools have been important to our success. Cloud-like
displays of information would complement our other advanced
features, such as geographic mapping and transcription
synchronisation.
The word clouds produced by TAPoR Tools, WMatrix and
Amazon are, for browsing, an improvement over tabular
statistical information. There is an opportunity for other
corpus data to be enhanced by using a cloud. Linguists often
use collocational information as a tool to examine language
use. Figure 3 demonstrates a typical corpus tool output: Figure 3. British National Corpus through interface developed
by Mark Davies, searching for ‘bank’, showing collocates
http://corpus.byu.edu/bnc/ (accessed 23 November 2007)
The data contained in the table lends itself to visualisation as
a cloud. As with the word cloud, the list of collocates can be
displayed alphabetically. Co-occurrence frequency, like word
frequency, can be mapped to font size. This would produce an
output visually similar to the word cloud. Instead of showing
all corpus words, they would be limited to those surrounding
the chosen node word.
Another valuable statistic obtainable via collocates is that of
collocational strength, the likelihood of two words co-occurring,
measured here by MI (Mutual Information). Accounting for
this extra dimension requires an additional visual cue to be
introduced, one which can convey the continuous data of an MI
score. This can be solved by varying the colour, or brightness
of the collocates forming the cloud. The end result is shown
in figure 4: Figure 4. Demonstration of collocate
cloud, showing node word ‘bank’
The collocate cloud inherits all the advantages of previous cloud
visualisations: a collocate, if known, can be quickly located due
to the alphabetical nature of the display. Frequently occurring
collocates stand out, as they are shown in a larger typeface,
with collocationally strong pairings highlighted using brighter formatting. Therefore bright, large collocates are likely to be
of interest, whereas dark, small collocates perhaps less so.
Hovering the mouse over a collocate will display statistical
information, co-occurrence frequency and MI score, as one
would fi nd from the tabular view.
The use of collocational data also presents additional
possibilities for interaction. A collocate can be clicked upon
to produce a new cloud, with the previous collocate as the
new node word. This gives endless possibilities for corpus
exploration and the investigation of different domains.
Occurrences of polysemy can be identifi ed and expanded
upon by following the different collocates. Particular instances
of usage are traditionally hidden from the user when viewing
aggregated data, such as the collocate cloud. The solution is to
allow the user to examine the underlying data by producing
an optional concordance for each node/collocate pairing
present. Additionally a KWIC concordance can be generated
by examining the node word, visualising the collocational
strength of the surrounding words. These concordance lines
can even by reordered on the basis of collocational strength,
in addition to the more traditional options of preceding or
succeeding words.
This visualisation may be appealing to members of the public,
or those seeking a more practical introduction to corpus
linguistics. In teaching use they not only provide analysis, but
from user feedback, also act as stimulation in creative writing.
Collocate searches across different corpora or document sets
may be visualised side by side, facilitating quick identifi cation
of differences.
While the collocate cloud is not a substitute for raw data, it
does provide a fast and convenient way to navigate language.
The ability to generate new clouds from existing collocates
extends this further. Both this iterative nature and the addition
of collocational strength information gives these collocate
clouds greater value for linguistic research than previous cloud
visualisations.