'Something that is interesting is interesting them': Using Text Mining and Visualizations to Aid Interpreting Repetition in Gertrude Stein's The Making of Americans

With The Making of Americans1, Stein’s goal was to
record “being” as it is manifested in repetition. Lauded
by some critics who thought Stein accomplished what T.S. Eliot
demanded of all writers—to make art, literature, and language
“new”—she was also criticized by others like Malcolm Cowley
who said Stein’s “experiments in grammar” made her novel
“one of the hardest books to read from beginning to end that
has ever been published.”2 More recent critics have attempted
to aid interpretation by charting the correspondence between
structures of repetition and the novel’s discussion of identity
and representation. Yet, the use of repetition in The Making of
Americans is far more complicated than manual practices or
traditional word-analysis programs (such as those that make
concordances or measure word-frequency occurrence) could
indicate. The text’s large size (almost 900 pages and 3183
paragraphs), its particular philosophical trajectory, and its
complex patterns of repetition make it a useful case study for
analyzing the interplay between the development of text mining
tools and the hermeneutics employed in interpreting literary
texts in general.
The reason that text mining procedures might be particularly
illuminating for Stein’s The Making of Americans is not that
these procedures make reading the text easier or because text
mining might “solve” a literary conundrum. In fact, such a
reductive approach would be very uninteresting to the literary
scholar. Many Stein critics argue that the “difficulty” the
repetition engenders in The Making of Americans is valuable
precisely because it is rooted in indeterminacy: that is, the text’s
use of repetition represents a postmodernist project that
challenges readerly subjectivity and deconstructs the role
language and the process of writing plays in determining
meaning. The more “difficult” the text is to read, the more it
meets its philosophical goals of making the reader question
acts of representation and interpretation in general. Thus the
text mining process is useful for literary analysis for three
reasons:
1. Text mining may be used to determine relationships (clusters
and patterns) in large bodies of data (in this case, the
confusing network of Stein’s repetitions ).
2. Text mining can be used to identify relationships within the
text that are not predetermined by meaning but are based
on structural elements of the content (such as repetitive
phrases or words).
3. Text mining also requires “subjective human evaluation”
as an essential part of the analytical process.3
Analyzing The Making of Americans has already provided rich
opportunities for thinking about both tool development and
processes of literary analysis. For example, initial analyses on
the text using the Data to Knowledge (D2K) application
environment for data mining4 (which was used in the nora
project5) has yielded clusters based on the existence of frequent
patterns. Instead of using single words for analyzing the text,
the features used were phrases (i.e. ngrams). Executing a
frequent pattern analysis algorithm6 produced a list of patterns
of the phrases co-occurring frequently in paragraphs. This
frequent pattern analysis generated thousands of patterns
because slight variations generated a new pattern. Because the
number of frequent patterns was so large, another algorithm
was applied that clustered these frequent patterns.7Although
the analytics that were used are sophisticated, the results are
not presented in a manner that makes them easy for scholars to understand. For example, one cluster has 85 frequent patterns
(see Figure 1). A scholar could piece clustered phrases back
together to form a longer repetitive phrase (“part of Gossols
where no other rich people were living”), but this is difficult,
and, in addition, the results do not indicate where the phrase
occurs in the document. The impact of Stein’s repetition, after
all, is always contingent on context, on what comes before and
after a particular string of letters, on what those strings (or
words) mean in the context of a sentence, a paragraph, and the
larger narrative or project according both to Stein and to the
scholar/user’s theoretical perspective.
Thus, this first attempt at producing clusters was only
marginally useful for scholarly study because the rules by which
the items are clustered are not readily apparent and the results
are detached from the context of the text. One method for
helping scholars understand the rules is by including the scholar
in identifying the features by which the frequent patterns are
determined. For instance, she could develop a “productive”
question about the text, a question that correlates to specific
linguistic and narrative features that can be extracted and
eventually mined. For example, a question might be: how do
linguistic and syntactic variations reflect, advance, or deter
from the progression of the text’s more abstract features such
as argument and narrative? Specifically, do changes in repetition
correspond to the novel’s evolving theories about identity and
representation? And how? Features which might be considered
include, but are not limited to, repetition across words and
phrases (including “trivial” and “motivated” recurrences);8
grammatical features (such as shifting perspectives, oscillating
verb-tenses, or dropped pronoun referents); syntactic features
(such as diagrammic versus run-on or fragmented sentences);
and narrative features (such as shifts between description or
“telling” and narrative or “showing”). More work needs to be
done to create the features described above for use in the
analytics. This paper will discuss our attempts to augment these
analyses.
In addition, in order to link these patterns back to the context,
it is essential to develop an environment that facilitates a
self-reflective, critical evaluation of the scholar’s theoretical
perspective and ultimately, her ability to influence how data is
culled, to assert what Ben Shneiderman has called “user
control.” Thus, the text’s features must be visualized in a
meaningful way so the scholar can “see” the text mining results
and engage with the techniques to improve the results. For
instance, in our first attempt, we created frequent patterns
without using stemming. After reviewing the list of results, we
realized that stemming was valuable to regulate pluralism, verb
tense, etc. Stemming was deployed. Our experience with the
norainterface has also demonstrated that providing easy access
to context within the text is important to understanding the
results of classification algorithms—a task which is relatively
easy when the entire text is a short poem (see Figure 2). A
scholar who is attempting to study repetition across the
900-page text, however, needs new ways to access the context
of each instance of repetition. She needs tools that take her
beyond searching tools which list the context of a single line
(see Figure 3 for results using “grep” on a file). To this end,
we have explored various visualization methods such as a
weighted centroid (Figure 4), a scatterplot (Figure 5), a heat
map (Figure 6), and a line graph (Figure 7).After our team
considered the advantages and disadvantages of each of these
tools, Anthony Don led the effort to design FeatureLens (Figure
8), a tool which allows the user to see features derived from
text mining results within the context of the original text. This
tool will serve as the main focus of this talk.
Ben Shneiderman calls the powerful combination of text mining
and visualization, “discovery” based because it empowers users
“to specify what they are seeking and what they find
interesting.”9 Gertrude Stein has this to say about the same
topic:
It is very likely that nearly every one has been very nearly certain
that something that is interesting is interesting them . . . The only
thing that is different from one time to another is what is seen and
what is seen depends upon how everybody is doing everything.
This makes the thing we are looking at very different and this
makes what those describe it make of it, it makes a composition,
it confuses, it shows, it is, it looks, it likes it as it is, and this makes
what is seen as it is seen.10
Text mining allows the literary scholar to “see” the text
“differently” by facilitating analytical approaches that chart
repetition across thousands of paragraphs and visualizations
that empower her to tweak the results, focus her search, and
ultimately (re)discover “something that is interesting.” This
paper will discuss how this process of analyzing repetition in
The Making of Americans has impacted the ways in which nora
has combined text mining and visualization applications.
Figure 1: A list of frequent patterns for one cluster Figure 2: With the nora interface, users can review the results of the data
mining, see which documents contain the feature words returned by the
algorithm, and see the location of the words in a selected document. ()
Figure 3: Identifying repetition as it emerges in a list of frequent pattern clusters
in a simple "grep" text file.
Figure 4: Using Text Arc, we are able to see where the word "Repeating" is
repeated throughout the entire second section of The Making of Americans.
The text of the section is represented by the lines in the outer ring. The word
"repeating" is situated in the circle according to where it appears most often
in the outer ring; the green lines represent lines in which the word appears;
and the orange lines point to word's occurrence in those lines. ()
Figure 5: Common phrases displayed on a scatterplot, with frequencies in
section 1 on the X axis and frequency in section 2 on the Y axis. We can see
than "men and women" and "I was saying" is a lot more common than any
other phrases, and used equally in both sections. ()
Figure 6: This heat map allows us to compare the frequency of phrases in five
sections. Each line is a different phrase. The red lines show when a phrase
occurs more than 100 times in a section. ()
Figure 7: This line graph shows frequent phrases compared across five sections.
Each line represents a different phrase. Here we can immediately see that "men
and women" appears almost as frequently in two sections of the text. () 1. Gertrude Stein, The Making of the Americas: Being a History of
a Family's Progress, (Normal, IL: Dalkey Archive Press, 1995).
First published by Contact Editions, Paris, 1925.
2. Malcolm Cowley, "Gertrude Stein, Writer or Word Scientist" The
Critical Response to Gertrude Stein, (Westport, CT: Greenwood
Press, 2000): 148.
3. This description of text mining is detailed further in Sholom M.
Weiss et al., Text Mining: Predictive Methods for Analyzing
Unstructured Information, (New York: Springer Science+Business
Media, Inc., 2005)
4. Developed by the Automated Learning Group (ALG) at the
National Center for Supercomputing Applications (NCSA),
alg.ncsa.uiuc.edu
5. The nora project (www.noraproject.org) is a Mellon-funded
collaborative (including computing, design, library science, andd
English departments at the University of Alberta; University of
Illinois, Urbana-Champaign; University of Maryland; University
of Nebraska; and the University of Virginia) which is developing
text mining and visualization software in order to "explore
significant patterns across large collections of full-text humanities
resources."
6. J. Pei, J. Han, R. Mao, "CLOSET: An Efficient Algorithm for
Mining Frequent Closed Itemsets (PDF)", Proceedings of the
2000 ACM-SIGMOD International Workshop on Data Mining
and Knowledge Discovery (DMKD'00), Dallas, TX, May 2000.
7. Xifeng Yan, Hong Cheng, Jiawei Han, and Dong Xin,
"Summarizing Itemset Patterns: A Profile-Based Approach,"
Proceedings of the 2005 International Conference on Knowledge
Discovery and Data Mining (KDD 05), Chicago, IL, August 2005.
8. "Trivial" and "motivated" recurrences are used by Calvin Brown
in his book Repetition in Zola's Novels, (Athens, GA: University
of Georgia Press, 1952). Examples of such recurrences include
"tags" (repeated descriptions such as "her face was worn, her
cheeks were thin," "her worn, thin, lined determined face," "her
lined, worn, thin, pale yellow face"), "key passages" (a relatively
long repetition of a fundamental idea) and "hammer word" (a
strong or exclusive obsession with any single idea).
9. Ben Shneiderman, "Inventing Discovery Tools: Combining
Information Visualization with Data Mining." Information
Visualization 1.1 (2002): 11.
10. Gertrude Stein, Composition as Explanation, (London: The
Hogarth Press, 1926). Accessed 2006-11-01. <http://www.
centerforbookculture.org/context/no8/st
ein.html>
Page