The Story of One: Developing text mining procedures and visualizations in the MONK project to re-read Gertrude Stein's The Making of Americans

Most critiques of The Making of Americas (Paris
1925) by Gertrude Stein contend that the text deconstructs
the role narrative plays in determining identity
by using indeterminacy to challenge readerly subjectivity.
The current perception of Making as a postmodern
text relies on the notion that there is a tension created
by frustrated expectations that result from the text’s progressive
disbandment of story and plot as the narrative
unweaves into seemingly chaotic, meaningless rounds
of repetitive words and phrases. Yet, a new perspective
that is facilitated by digital tools and based on the
highly structured nature of the text suggests that these
instabilities can be resolved by the same seemingly nonsensical,
non-narrative structures. Seeing the manner in
which the structure of the text makes meaning in conversation
with narrative alleviates perceived instabilities in
the discourse. The discourse about identity formation is
engaged—not dissolved in indeterminacy—to the extent
that the reader can read the composition.
One method for reading the composition of the text without
relying on what becomes a non-existent framework
based on plot is to view the progression of words according
to a different framework, a framework that relies
on comparative associations based on word usage.
Using WordHoardi, we compared word usage between
texts and text parts by calculating the log-likelihood ratio,
which describes the size and significance of the difference
between word frequencies in a base text versus
a reference text.ii In this analysis, we measured word
usage in The Making of Americans in comparison to two
different sets of reference texts with more traditional
narrative structures: (a) a set of 19th century novels written
by Jane Austen, Charles Dickens, George Eliot, and
George Meredith;iii (b) between the first and second half
of Making, which it has been argued also represents different
narrative trends (Clement 2008). Visualizing this
information in Wordleiv—a word cloud application (Wattenberg
& Viegas, 2008)—is useful primarily because
it provides a visual overview of word frequencies that
is easy to understand and to publish for reference. The
Wordle application facilitates this kind of analysis by
visualizing the list of words in a cloud that maximizes
the space utilization on a computer screen by sizing the
words by their relative frequencies. The more frequently
a word occurs in a particular text (relative to another
text) the larger the word appears. In the set of visualizations
that accompany this discussion,v each cloud
serves to visualize words that are more or less frequent
in any given comparison (see examples in Fig. 1). What
becomes immediately evident in comparing these visualizations
is the prominence of a particular word that
consistently scores a high value in terms of discrepancy
between Making and the reference texts: one. The word one appears consistently across every cloud
that marks the words that are more common in Making
and less common in the sample of nineteenth century
texts and words that are more common in the second half
rather than the first half of the text. We then compared
the relative frequencies of multiple pronouns, which revealed
that the frequency of one surges by the end of the
text (Fig. 2). What is most interesting about this graph is
that the high frequency of one is the result of the confusion
accomplished by the word’s schizophrenic nature.
Words here are represented according to occurrence, not
to type of occurrence. Thus, the word one—unlike he,
she, I, we, or even you or it—which plays many positions
in the text, in the role of a pronoun or an adjective and in
the subject or object position, surges in frequency.
To better understand how the word one is used in the
text, we created another set of visualizations prepared
using a prototype we developed called PosViz.vi Pos-
Viz allowed us to compare word usage based on parts of
speech in individual chapters from Making to the whole
text. These comparisons allowed us to isolate and analyze
words used more and less frequently throughout the
course of the novel and to measure how and if these patterns
change within the text itself. In these visualizations,
each part of speech for each word is treated as a separate
word instance and each instance is placed according to
its appearance in the text, color-coded according to its
part of speech, and sized according to its overall frequency.
Thus, by using PosViz, the progression of the manner
in which the word one is used in terms of different
parts of speech is documented, allowing us to see that the
use of one appears to change as the text progresses. For
example, a relatively small one appears three times in
the chapter 1 cloud (Fig. 3). This visualization indicates
that the occurrence of the word one has little variance
in terms of how it is used (its part of speech) and occurs
relatively infrequently in occurrences that are localized
to the beginning paragraphs of the chapter.vii By chapter
9, however, one dominates the discourse both in terms of
its frequency and in terms of its multiple uses (Fig. 4).
By identifying the manner in which word usage changes
in correlation to the presence and absence of narrative
both in comparison to other novels and within the text
itself, these comparisons enable a new perspective on the
meaning-making processes of the text’s composition. For
example, these visualizations illustrate the nature of the
word one as it is used to heighten the word’s propensity
for different reading possibilities. This lends to a reading
in which one may represent a singular subject position or
multiple subject positions at once. With this information,
a further argument can be made that the discourse about
identity formation is engaged in this multiplicity, not dissolved
in indeterminacy. Thus, employing composition
in her representation of identity formation in The Making
of Americans becomes the method by which Stein seeks
to represent identity, but if and how the reader is able
to recognize and interpret this endeavor is predicated by
her ability to see it.
This work with The Making of Americans is part of research
and development within the MONK (Metadata
Offer New Knowledge) project, a Mellon-funded collaborative
seeking to develop text mining and visualization
software in order to explore patterns across large-scale
text collections. Stein’s text was a productive text for
analysis during the beginning phases of the MONK project
since its many and complicated repetitions could be
processed and visualized.viii This presentation focuses
on how the process of determining decision criteria for
text mining led to the discovery that various textual features
(n-grams, parts-of-speech, and log-likelihood ratios)
and various visualizations (FeatureLens, Spotfire,
Wordle, and PosViz) ultimately facilitated an iterative
discovery process and a new reading of Gertrude Stein’s
The Making of Americans. Notes
iPlease see http://wordhoard.northwestern.edu/userman/.
iiThis analysis is based on Dunning’s log-likelihood
analysis. Please see http://wordhoard.northwestern.edu/
userman/analysis-comparewords.html#loglike.
iiiThe books used in this study are those available in the
WordHoard application. They are listed at http://terpconnect.
umd.edu/~tclement/less-more.htm. These authors
were chosen because Stein repeatedly compares her text
to their novels. See Stein 1990, p. 506.
ivPlease see http://wordle.net/.
vThese visualizations are pictured in an online appendix
entitled “Visual Comparisons of Gertrude Stein’s The
Making of Americans using WordHoard, Wordle, and
PosViz” at http://terpconnect.umd.edu/~tclement/lessmore.
htm. In these slides, the comparisons have been
visualized with four sets of data for each data set, all of
which are set in comparison to the base text The Making
of Americans and include and exclude ‘common words’
such as articles, conjunctions, and pronouns. The full
list of these ‘stop-words’ is unavailable, but the creator
Jonathan Feinberg has indicated in an email that ‘I have
modified them by hand over time. The English one came
from the Snowball stemmer project’ at http://snowball.
tartarus.org/algorithms/english/stop.txt (personal correspondence).
viCurrently, PosViz does not have a web presence.
viiThe analysis program used to label these uses is part
of the SEASR (Software Environment for the Advancement
of Scholarly Research) analytic routines (http://
seasr.org/). Though imperfect, the system is consistent—
it labels the same behaviors the same way each time.
Thus each occurrence represents the perception of a different
use of the word one. viiiThis work is published in two articles: Don et al., 2007
and Clement, 2008.
References
Clement, T. (2008). “‘A thing not beginning or ending’:
Using Digital Tools to Distant-Read Gertrude Stein’s
The Making of Americans.” Literary and Linguistic
Computing, 23.3: 361-382.
Don, A., Zheleva, E., Gregory, M., Tarkan, S., Auvil,
L., Clement, T., Shneiderman, B., Plaisant, C.
(2007). “Discovering interesting usage patterns in text
collections: integrating text mining with visualization.”
Proceedings of the sixteenth ACM conference on Conference
on information and knowledge management, pp.
213-222.
Stein, G. (1990). “Transatlantic Interview 1946.” In The
Gender of Modernism. Bonnie Kime Scott and Mary
Lynn Broe (eds). Bloomington: Indiana University
Press, pp. 502-516.
Wattenberg, M. and Viegas, F. (2008). “Tag clouds
and the case for vernacular visualization.” Interactions,
15.4: 49-52.
Stein, G. (1995). The Making of Americans: Being a
History of a Family’s Progress. Normal, IL: Dalkey Archive
Press.
Weiss, Sholom M. et al. (2005). Text Mining: Predictive
Methods for Analyzing Unstructured Information. New
York: Springer, pp. 85-86.