introducing nora: poetry, pattern recognition, and provocation

1
introducing nora poetry, pattern recognition,
and provocation

• Matthew Kirschenbaum
• University of Maryland
• mgk_at_umd.edu

2
the nora project

• Funded for two years, 600k, by the Andrew W.
  Mellon foundation
• Were one year in
• PI John Unsworth (Dean, GSLIS, Illinois)
• Participating researchers from Illinois (GSLIS,
  NCSA), Georgia (English), Maryland (MITH, HCIL,
  English), Virginia (IATH, CS, English), Alberta
  (Humanities Computing)
• Fields of expertise include literary studies,
  library science, computer science
• www.noraproject.org

3
technologies and heuristics

• Digital libraries
• Text mining
• Machine learning
• Iteration and play
• Visualization
• Provocation and anomaly

4
old school

• According to Aristotle's commendable formula . .
  . the beautiful is defined as that which the eye
  can easily embrace in its entirety and which can
  be surveyed as a whole. The tragedy of King
  Oedipus may well arouse pity and fear, but
  according to The Poetics it is beautiful because
  it fulfills the temporalized optical requirement
  of having a beginning, a middle, and an end.
  Perception of its form is not resisted by
  boundlessness. Thus, long before Baumgarten's
  modern foundation of the concept and the subject
  matter of Aesthetics, and longer still before the
  term arose which will have guided my commentary,
  Aesthetics begins as "pattern recognition.
• --Friedrich Kittler, "The World of the
  Symbolic--A World of the Machine"

5
digital libraries

• A lot of us have spent the last 10 to 20 years
  digitizing large piles of texts and other media
• We can browse these online archives and
  collections, and search them in reasonably useful
  ways. But what else can we do?
• nora starts with 5 GB of 18th and 19th century
  British and American literature contributed from
  about a dozen different repositories and
  collections

6
text mining

• The semi-automated discovery of trends and
  patterns across very large datasets (Hearst 3).
• Don Swansons association of magnesium deficiency
  with migraine headaches by mining bio-medical
  literature in the 1980s
• A new, potentially plausible medical hypothesis
  was derived from a combination of text fragments
  and the explorers medical expertise (Hearst 6).
• Everyday Uses of This
• Classification
• Matching
• Clustering
• Prediction

7
classification (sorting)
8
matching (More like this . . .)
9
clustering
10
prediction (machine learning)
11
i. emily dickinson, hot or not?

• Prediction can the machine be taught to
  identify patterns of erotic language in the
  corpus of correspondencesome 200
  lettersexchanged between Emily Dickinson and
  Susan Huntington Dickinson (Emilys
  sister-in-law)?
• We remind her we love her - Unimportant fact,
  though Dante did'nt think so, nor Swift, nor
  Mirabeau.

12
what we did

• Two members of the project team with expertise in
  Dickinsons writings, Martha Nell Smith and Tanya
  Clement, sat down with the corpus and labeled
  each text hot or not. Yes, this was to an extent
  arbitrary.
• These evaluations were passed to a data mining
  expert (Bei Yu) who subjected the corpus to a
  kind of predictive analysis known as Naïve
  Bayesian classification.

13
the black box

• Bayes is Thomas Bayes (1702-1761), the British
  mathematician and minister
• Bayesian probability is the domain of probability
  that deals with non-quantifiable events not
  whether a coin will land heads or tails for
  instance, but rather the percentage of people who
  believe the coin might land on its side also
  known as subjective probability
• Our Bayesian classification is naïve because it
  deliberately does not consider relationships and
  dependencies between words we might instinctively
  think go togetherkiss and lips, for example.
  The algorithm merely establishes the presence or
  absence of one or more words, and takes their
  presence or absence into account when assigning a
  probability value to the overall text. This is
  the kind of thing computers are very good at, and
  naïve Bayes has been proven surprisingly reliable
  in a number of different text classification
  domains.
• The math for this is
• Which is the probability that a given document D
  belongs to a given class C.

14
raw output

• Some 4000 featuresdiscrete word typesextracted
  and ranked
• CLASS_PROB_Hot CLASS_PROB_Not RATIO FEATURE
  double double double String 7.140671223094971E-4
  7.871536523929471E-5 2.2051385922805053 mine
  7.140671223094971E-4 7.871536523929471E-5
  2.2051385922805053 must 6.120575334081404E-4
  7.871536523929471E-5 2.0509879124532464 Bud
  6.120575334081404E-4 7.871536523929471E-5
  2.0509879124532464 Woman 6.120575334081404E-4
  7.871536523929471E-5 2.0509879124532464 Vinnie
  6.120575334081404E-4 7.871536523929471E-5
  2.0509879124532464 joy 6.120575334081404E-4
  7.871536523929471E-5 2.0509879124532464 Thee
  5.100479445067837E-4 7.871536523929471E-5
  1.8686663556592924 write 5.100479445067837E-4
  7.871536523929471E-5 1.8686663556592924 Eden
• 5.100479445067837E-4 7.871536523929471E-5
  1.8686663556592924 luxury 5.100479445067837E-4
  7.871536523929471E-5 1.8686663556592924 Alps
  5.100479445067837E-4 7.871536523929471E-5
  1.8686663556592924 weary 5.100479445067837E-4
  7.871536523929471E-5 1.8686663556592924 sick
  5.100479445067837E-4 7.871536523929471E-5
  1.8686663556592924 Hour 5.100479445067837E-4
  7.871536523929471E-5 1.8686663556592924 Dream
  0.009384882178924818 0.0014955919395465995
  1.8365780411077903 you 0.001836172600224421
  3.1486146095717883E-4 1.7633058400014656 Susie
  4.080383556054269E-4 7.871536523929471E-5
  1.6455228043450818 remember 4.080383556054269E-4
  7.871536523929471E-5 1.6455228043450818 garden
  4.080383556054269E-4 7.871536523929471E-5
  1.6455228043450818 dream 4.080383556054269E-4
  7.871536523929471E-5 1.6455228043450818 cant
  4.080383556054269E-4 7.871536523929471E-5
  1.6455228043450818 always 4.080383556054269E-4
  7.871536523929471E-5 1.6455228043450818 loved
  4.080383556054269E-4 7.871536523929471E-5
  1.6455228043450818 lives

15
some results
Slide credit Bei Yu
16
ii. deduction

• Can we decide what humanists might want to do
  with text mining technology by using deductive
  methods?
• Compare verbs from scholarly prose in ALH and ELH
  to usage of these same verbs in the American
  National Corpus (drawn from the NY Times and
  other newspapers)

17
holding up the mirror

• totalizing
• reprinting
• mediating
• slaveholding
• privileging
• absenting
• misreading
• desiring
• seafaring
• fetishizing

• What would it mean to teach a machine to
  recognize the way in which a text privileges,
  fetishizes, or totalizes?
• Wiser heads than ours are working on it. Dominic
  Widdows, Geometry and Meaning (Stanford CSLI
  Publications, 2004).

18
iii. it always ends in tears reading sentiment

• Can we teach a machine to recognize and read
  sentimental literature?
• Stage 1. Evaluate the use of text-mining on a
  small set of core" sentimental novels. We will
  label a subset of the chapters (the training set)
  with a a score indicating a level of
  sentimentalism, and then see how text mining
  classifies the remaining chapters from those
  novels. (Texts Charlotte, Uncle Toms Cabin,
  Incidents in the Life of a Slave Girl.)
• Stage 2. Two more novels will be added to the set
  to evaluate how well the processes work when more
  texts by the same authors are added to the set of
  works studied. (Texts added Charlottes
  Daughter, The Ministers Wooing.)
• Stage 3. The texts added will be those that
  scholars recognize as exhibiting sentimentalism,
  though some may not be as consistently
  sentimental chapter-by-chapter as the "core" set
  used earlier. In this experiment there will be
  more focus will be on gaining insights on
  sentimentalism and these novels than in previous
  experiments. (Some likely texts Clotelle, The
  Lamplighter, The Coquette, Hobomok. )
• Stage 4. Using a text-mining model that was
  developed to identify chapters with strong
  sentimentalism, use text-mining on a set of works
  that are considered by scholars not to be wholly
  sentimental or just not sentimental at all. This
  may identify parts of texts that contain aspects
  of sentimentalism, or common word-use that is
  sentimental in one novel but not another. In this
  experiment there will be a strong focus will be
  on gaining insights on sentimentalism and these
  novels. (Texts Moby-Dick, The Scarlet Letter,
  The Blithedale Romance, Irvings Sketchbook,
  Narrative of the Life of Frederick Douglass.)

19
a general use scenario

• User begins with a set of documents, obtained
  from one or more federated digital library
  collections.
• The computer puts up a random list of ten items
  in your subset, and you rank each one, either on
  a yes/no scale, or on a 1-5 scale.  We warn you
  that if yes/no, less training will be required
  if 1-5, more training will be required.  I mark
  the ones that are like what I'm looking for, and
  the ones that are not.  Each subsequent retrieval
  of 10 items builds on the last, inasmuch it takes
  your last ranking into account as it evaluates
  semantic distance from your best examples---or
  possibly employs some other similarity measure.
   This positive/negative ranking is the basis for
  the next round....
• Having been told that a particular set is rich in
  the feature ofinterest, and another set is weak
  in that, our software brings up iterative guesses
  about what's of interest, and invites the user to
  rank known items.
• Arriving at what seems like the optimal set of
  "things like what I'm interested in," the user
  pushes a button to ask what things seem to
  predict this set--what are the characteristics
  that are highly correlated with this feature?

20
the architecture
21
visualization and graphesis
22
provocation and anomaly

• My machine doesn't just analyze and spit out
  results. All it says is, Here is strangeness.
  Im the one who gets to look.
• --Randall McLeod, working at his collator

23
www.noraproject.org
24
acknowledgements

• Members of the nora project team whose work
  contributed most directly to the results
  presented here include Loretta Auvil, Tanya
  Clement, Tom Horton, Greg Lord, Catherine
  Plaisant, Steve Ramsay, James Rose, Martha Nell
  Smith, Sara Steger, Kristen Taylor, John
  Unsworth, and Bei Yu

25
the introduction

• This is Nora, Stella says, stepping softly
  past Cayce to lay her hands on the shawl-draped
  shoulders of the figure in the chair before the
  screen. Noras right hand pauses. Still resting
  on the mouse, though Cayce senses this has
  nothing to do with her sisters touch, or the
  presence of a stranger.
• --William Gibson, Pattern Recognition