Towards a rhizomatic narrative

Abstract

We present Crossreads, a manner to deconstruct linear narrative text in order to read text in multiple orders. This is an ongoing project aims to study data multiplicity, as well as textual visualization interfaces. The process starts with the selection of a text, which is later segmented into small blocks, and the textual similarity among them is calculated, forming a network data set. Finally, a web interface allows the user to explore and read through the created network of text.

Full text

Crossreads
Towards a rhizomatic narrative
Jaume Nualart
University of Canberra (AU)
University of Barcelona (CAT)
National ICT Australia, NICTA (AU)
jaume.nualart@canberra.edu.au
Gabriela Ferraro
National ICT Australia, NICTA (AU)
Australian National University (AU)
gabriela.ferraro@nicta.com.au
ABSTRACT
We present Crossreads, a manner to deconstruct linear nar-
rative text in order to read text in multiple orders. This is
an ongoing project aims to study data multiplicity, as well
as textual visualization interfaces. The process starts with
the selection of a text, which is later segmented into small
blocks, and the textual similarity among them is calculated,
forming a network data set. Finally, a web interface allows
the user to explore and read through the created network of
text.
Categories and Subject Descriptors
H,5,4 [information Interfaces and Presentation]: Hy-
pertext/Hypermedia: navigation
General Terms
Experimentation
Keywords
narrative, deconstruction, data multiplicity, visualization.
1. INTRODUCTION
Inspired by works such as Rhizome [4] we present an ongoing
project in the domain of information seeking and discovery
called Crossreads. This project proposes an experimental
way of reading texts, as an alternative to traditional linear
reading. I proposes to break the initial narrative line of a
text by segmenting it into smaller parts. Then, the text is
reordered according to similarity scores, which finally offers
the reader multiple paths to read the text. The aim of this
project is to explore and study the effects when a reader
processes fragmented information, as well as to analyze user
activity and support reader’s exploration with visualization
techniques: the interaction text-reader, the text as a collec-
tion of segments, the most popular reading paths, and so
on. At this stage of the project we cannot tell accurately
what the benefits of the crossreading are unclear. However,
according to some authors, learners naturally make connec-
tions between pieces of knowledge, and they are better able
to retrieve and apply their knowledge when those connec-
tions are accurate and meaningful [1].
Crossreads’ outputs are a network data set where nodes are
segments of text, and an interface that supports nonlinear
reading.
This project is part of a practice-led PhD project that has
three main focuses: the study of text visualization approaches,
the multidisciplinary of the field, and the process of design
and development tools. The presented approach, Cross-
reads, is one of the artefacts created as part of the PhD.
The whole project includes a proposed new classification of
text visualization tools [10], a visualization tool for single
texts [9], and two visualization tools to explore and overview
collections of texts [5], [8].
In the following sections, we discuss related work and influ-
ences. Then, we present the experiments done so far. Fi-
nally, we present some conclusions and directions for future
work.
2. RELATED WORK
Several works in the past have explored the possibilities
of breaking the linearity of a text. As mentioned in the
introduction, the philosophers Deleuze and Guattari have
described the rhizomatic structure of knowledge, which in-
spires this project too: “In a book, as in all things, there are
lines of articulation, segmentarity, strata and territories; but
also lines of flight, movement, deterritorialization and de-
stratification”. In the novel Hopscotch by J. Cort˜
A
azar [3],
the author proposes two reading order for the chapters; the
text starts with: “In its own way this book is many books,
but mostly it’s two books”. The Project Xanadu from 1960
[12] is considered the first hypertext project in the digital
era, and it was a visionary definition of standards for the
WWW that were mostly not included in the standard pro-
tocols. One of Xanadu’s rules states: ”Every document can
consist of any number of parts each of which may be of any
data type.”. The open Xanadu pro ject is accessible and like
Crossreads, it encourages nonlinear navigation of text. The
aim of Xanadu’s demo is to demonstrate the possibilities of
hypertext.
These are the main examples that make us to investigate the
effects of reading in alternative ways in combination with

normal reading.
3. EXPERIMENTS
Currently, two versions of Crossreads have been developed
(i.e., Version I and II). Version I is part of an exhibition at
Museum of Contemporary Art of Barcelona (MACBA), with
texts in Catalan and Spanish by the artist Eugeni Bonet
[7]. Version II uses texts in English by Domenico Quaranta
about media art, compiled in the book “In Your Computer”
[11], and it is accessible on line [6]. In both cases, the texts
used are licensed under Creative Commons.
The creation of crossreads implies two different tasks: (i)
data preparation and analysis. and (ii) interface design.
3.1 Data preparation and analysis
During the data preparation and analysis stage, three main
steps have been identified: data set selection, data segmen-
tation, and similarity calculus.
3.1.1 Data set selection
So far, we have experimented using data sets from a sin-
gle author and, more research should be done in order to
propose text collections from multiple authors, topics, lan-
guages and other criteria.
The original data used for this experiment, in the two ver-
sions, have a particular feature: the data, i.e. the texts, are
document collections that contained opinion and critic arti-
cles, compiled in books. We designed Crossreads respecting
the original documents. The interface also enables linear
reading of each document of the collection. We do not an-
ticipate however that there will be any significant design
problems if the original data come from a single document.
3.1.2 Data segmentation
We have experimented with two segmentation approaches,
each with different benefits. In both versions of Crossreads
each document was divided into segments, such that each
segment consisted of one or more paragraphs. A segment
length was about seven hundred characters in total; which
equates to an average of one minute of reading for an adult
[13]. In Version I, segmentation was machine produced. In
Version II, segmentation was performed manually. While the
method used in Version I was fast and capable of processing
large collections, the method applied in Version II method
allowed for a greater quality segmentation.
The reason for these two approaches is that: the segmenta-
tion task is very subjective. A human expert could add a
personal view to the segmentation (Version II). A machine
produced segmentation (Version I) can accomplish well this
task in terms of size of each segment, but it cannot be ex-
pected the richness of an expert. We wanted to compare
both methods as part of the initial experimentation with
the intention that it will be further validated by a user eval-
uation test.
3.1.3 Data similarity
To develop the similarity calculus between segment neces-
sary to create the Crossreads network, we used the following
off-the-shelf Natural Language Processing tools and tech-
niques:
•Tokenization: words in the segments are separated by
white space and punctuation characters.
•Stop word removal: standard stop word removal.
•Named Entity Recognition: identification and classifi-
cation of Named Entities (NE) in each segment. We
applied the OpenNLP Named Entity recogniser [2],
which distilled four types of entities, Person, Location,
Organization and Others.
•Similarity Calculus between segments.
The similarity between pairs of segments was calculated as
the sum of the following factors,
Sim(i, j ) = T okSim +EntitySim +N ESim/3
where TokSim is the token cosine similarity between seg-
ments, which is a common vector based similarity measure.
To calculate the similarity, the tokens of each segment are
transformed into vectors and then the Euclidean cosine is
used to determine the similarity between pairs of vectors;
EntitySim is the sum of the NEs in each segment, normal-
ized by the number of tokens in both segments; and NESim
is the cosine similarity between NE. During this process, the
similarity between different NE types (Person, Location, Or-
ganization and others) is calculated separately.
The similarity between the segments was calculated as fol-
lows. First, an arbitrary segment iwas chosen and used to
calculate the similarity between the segment iand the entire
segment collection. Second, the segment with the highest
similarity value score is set as the maximum similar seg-
ment of i. Since linear reading of a documents is enabled,
in each iteration we decided to skip links to segments of the
same document as segment i. Finally, we applied different
constraints to Version I and II, were as follows:
Version I: In the Crossreads network, each segment is linked
to its most similar segment. The drawback of this ap-
proach is that links will have a wide range of similarity
scores, since in each iteration, the number of segments
to compare with is smaller, and the possibility of find-
ing a segment with a high similarity score decreases.
However, the benefit is that there will not be any or-
phan segments, i.e. all segments link to other seg-
ments, so the reader will always have the possibility of
some crossreading.
Version II: Each segment is linked to its most similar pair-
ing. To avoid repetition of pairs, segments that have
already been set as a maximum similarity segment dur-
ing ten iterations are skipped. After ten iterations, the
skipped segments are used again in the similarity cal-
culus.
Again, both methods will need to be evaluated by users.

3.2 Interface design and visualization
The interface has been designed to optimal user experience.
It allows linear reading of the texts in combination with
crossreading. A reader can choose any text within the col-
lection to read. The collection of texts is presented in a
time-line, and in a flat list with text-category filters. Both
versions share a similar interface, with slight differences ac-
cording to the different methods of data analysis and authors
decisions.
In both cases there is a design principle: vertical naviga-
tion for linear reading of documents —using standard up
and down arrows images— , and horizontal navigation for
crossreading —using specially designed left and right arrows
images.
Version I of the interface was developed for the purposes
of a museum exhibition. Accordingly, the interface was in-
fluenced by a team of experts from MACBA including pro-
ducers, curators and art historians. In this version of Cross-
reads, when the user reads a segment of text, they can choice
between two links, one left and the other right. With the
right link, the reader goes to its most similar segment. With
the left link, the reader goes to the second most similar seg-
ment. Thus, both links offered the Crossreads experience.
In the links, it is announce the title of the document the links
goes to, proving the user with some context before following
the links. Version II of the interface has evolved such that
it offers link nuances. For instace, with the right link, the
reader goes to its most similar segment. The link context is
also shown as in Version I. Furthermore, the quality of the
link is represented with an icon, which shows the similarity
score between the current segment and the segment in the
right link, as well as the token and entities similarity scores.
With the left link, the reader jumps to a random segment
of the collection.
4. CONCLUSIONS AND FUTURE WORK
Crossreads proposes a novel way to explore a text collection,
based on text segmentation, the textual similarity between
the segmented pieces of texts, and a reader interface. For
future work, a user evaluation is planned in order to assess:
(i) the impact of the human and the automatic segmentation
approaches in the crossreads experience, (ii) how the simi-
larity among segments is interpreted by readers, and (iii)
the effect of crossreading in the learning process.
Furthermore, future work will focus in discussing the condi-
tions that a text must accomplish in order to suits the cross-
reading technique, in particular whether crossreadings is
suitable for one or multiple authors, one or multiple genres,
monolingual and/or multilingual collections, just to mention
a few variables.
5. REFERENCES
[1] Susan A Ambrose, Michael W Bridges, Michele
DiPietro, Marsha C Lovett, and Marie K Norman.
How learning works: Seven research-based principles
for smart teaching. John Wiley & Sons, 2010.
[2] Jason Baldridge. The opennlp project. URL:
http://opennlp. apache. org/index. html,(accessed 2
February 2012), 2005.
[3] Julio Cort´azar. Hopscotch (rayuela). New York:
Pantheon, 1966.
[4] Gilles Deleuze and F´elix Guattari. Introduction:
rhizome. A thousand plateaus: Capitalism and
schizophrenia, pages 3–25, 1987.
[5] Jaume Nualart. Area, 2013. [Accessed: 2014-07-27.
(Archived by WebCite at
http://www.webcitation.org/6RN89Vogt)].
[6] Jaume Nualart, Sonia Lopez, Gabriela Ferraro.
Crossreads and D. Quaranta, 2014.
[7] Jaume Nualart, Sonia Lopez, Gabriela Ferraro.
Crossreads at MACBA, 2014. [Accessed: 2014-07-27.
(Archived by WebCite at
http://www.webcitation.org/6RN8iHR48)].
[8] Jaume Nualart, Wray Buntine. Visference, 2013.
[Accessed: 2014-07-27. (Archived by WebCite at
http://www.webcitation.org/6RN7kqO7J)].
[9] J. Nualart and M Perez-Montoro. Texty, a
visualization tool to aid selection of texts from search
outputs. Information Research, 18(2), jun 2013.
[10] J Nualart, Mario P´erez-Montoro, and Mitchell
Whitelan. How we draw texts: a review of approaches
to text visualization and exploration. El Profesional
de la Informaci´on, 2014, vol. 23, num. 3, p. 221-235,
2014.
[11] Domenico Quaranta. In Your Computer. Lulu. com,
2011.
[12] Ted Nelson, et al. Project Xanadu, 1960. [Accessed:
2014-07-26. (Archived by WebCite at
http://www.webcitation.org/6RLo0HzFo)].
[13] James R Williams. Guidelines for the use of
multimedia in instruction. In Proceedings of the
Human Factors and Ergonomics Society Annual
Meeting, volume 42, pages 1447–1451. SAGE
Publications, 1998.