1 of 41

Suspense...

Chelsea Davis, Abigail Droge, Tasha Eccles, Laura Eidem, Morgan Frank, Erik Fredner, J.D. Porter, Andrew Shephard, Oleg Sobchuk, Hannah Walser, and Mark Algee-Hewitt

2 of 41

Key Questions

  • How does narrative suspense work?
  • What are its characteristics across genre and time?
  • “Where” is suspense? In structural features? In the reader’s affect? In semantics?
  • How might we be able to detect suspense in a text through digital means?

3 of 41

Recap from last time

  • Collaboratively generated corpora:
    • 213 suspenseful texts
    • 106 unsuspenseful texts
  • Through tagging (rating the degree of suspense felt paragraph by paragraph in) 14 short stories, we generated a list of most distinctive words (MDWs) for suspense and unsuspense

4 of 41

Native Son MDWs (April 2014)

5 of 41

Tracking features in suspenseful moments

6 of 41

Overview

  • Tagging: short stories and novels
  • Generating MDWs from tags
  • Semantic fields → topic models
  • Creating our own suspense detection model
  • Looking forward

7 of 41

Tagging Basics

  • Tagging short stories: suspense scale (0-10)
  • High degree of correlation among our tags
  • Tagging novels: non comprehensive, finding very high suspense, very low suspense
  • So far we’ve tagged around 40 novels

8 of 41

Our tags line up

9 of 41

Novel Tagging Statistics

10 of 41

11 of 41

Suspense MDWs

Unsuspense MDWs

Short story

Novel corpus

Novel tags

Short story

Novel corpus

Novel tags

upon

lorenzo

cairo

governor

utopia

colonel

eyes

digger

tom

masters

amelia

major

thought

carlyle

chamber

rhoda

barton

pyramid

saw

gran

temple

robert

ferris

dinner

horror

geoffrey

breath

loved

edith

discussed

sound

randy

fast

married

emerson

whoever

half

ryan

instant

dinner

sax

peter

dead

monroe

stranger

pretty

tuileries

inherent

hope

vic

van

lived

cecil

presumably

descent

greenwood

shoot

liked

dublin

relations

12 of 41

How closely do our three MDW groups track suspense in a novel?

13 of 41

14 of 41

15 of 41

Topic Modeling

  • What is a topic?
    • Most simply: a set of correlated words returned from a statistical tool
    • We culled and refined the “raw” topics generated by our model
  • Topicity
    • How distinct is a topic from other topics?
    • Network of correlation
  • Using topic model to generate new semantic fields
    • Cull words by posterior probability of words in topic (>0.5%)
    • Charting percentage of topic words that show up more often than expected in a moving window
  • Charting topic movement across narrative, rather than labeling texts

16 of 41

Topics vs. Semantic Fields

Violence (hand-generated)

murder

poison

strangle

assassin

rifle

assault

deadly

wounded

torture [... 55 total items]

Assault (topic modeled)

back

ran

toward

shouted

head

screamed

began

around

feet [... 36 total items]

17 of 41

Example Topics

  • Interior Spaces
    • door, room, hall, opened, open, window, stairs, steps, closed, entered
  • Polite Women
    • miss, lady, ladies, young, mr, dear, kind, house, sister, woman, daughter
  • Speech Verbs
    • said, replied, cried, will, answered, exclaimed, continued, returned, shall
  • Food
    • eat, food, coffee, kitchen, table, bread, milk, breakfast, good, hot

18 of 41

19 of 41

20 of 41

21 of 41

22 of 41

23 of 41

Suspense Detector

  • In the interest of testing our newly refined topic models and revised tagging system, we went back and tagged some more novels.
  • Overall, these experiments were a success. The moments that we in the group tagged as suspenseful generally corresponded to spikes for High Suspense MDWs.
  • In this regard, Morgan’s tagging of The Maltese Falcon yielded a particularly interesting result.

24 of 41

Maltese Falcon Results

25 of 41

Negative Presence and Suspense

  • Image here: Politesse in Dracula (1311Stoker1897.pdf)

26 of 41

Neural Network

Given the subjective nature of our identification/tagging process, we turned to the best computational model to duplicate this subjective process: a Neural Network

Our features seem to correlate with suspense, BUT suspense is complex

Could we use an equally complex model to combine our features and train it to recognize the same suspenseful passages as we did?

27 of 41

Neural Net: Schematic Diagram

Output: sections we identified as suspenseful or unsuspensful through our tagging experiment

Input: All topic model fields, MDW fields (except those extracted from tags) and age of acquisition scores - all scored in .5% slices of each text

28 of 41

neural 3

Neural Network: Success Rate

29 of 41

Neural Network: Feature Loadings

30 of 41

Neural Network: Feature Loadings

31 of 41

Automatically Tagged Text

<highSuspense3> the coroner did acquit me of all blame he had even complimented her on her presence of mind and courage she remembered for an inquest it couldnt have gone better and mrs hamilton had been kindness itself to her only hugo but she wouldnt think of hugo suddenly in spite of the heat in the carriage she shivered and wished she wasnt going to the sea a picture rose clearly before her mind cyrils head bobbing up and down swimming to the rock up and down up and down and herself swimming in easy practised strokes after him cleaving her way through the water but knowing only too surely that she wouldnt be in time the sea its deep warm bluemornings spent lying out on the sands hugo hugo who had said he loved her she must not think of hugo she opened her eyes and frowned across at the man opposite her a tall man with a brown face light eyes set rather close together and an arrogant almost cruel mouth she thought to herself i bet hes been to some interesting parts of the world and seen some interesting things iii philip lombard summing up the girl opposite in a mere flash of his quick moving eyes thought to himself quite attractive a bit school mistressy perhaps a cool customer he should imagine and one who could hold her own in love or war hed rather like to take her on he frowned no cut out all that kind of stuff this was business hed got to keep his mind on the job what exactly was up he wondered that little jew </highSuspense3>

32 of 41

Automatically Tagged Text

Agatha Christie: And Then There Were None - Tagged by Neural Network

33 of 41

Automatically Tagged Text

William Beckford: Vathek - Tagged by Neural Network

34 of 41

Automatically Tagged Text

William Gibson: Neuromancer - Tagged by Neural Network

35 of 41

Unsuspenseful Texts

Carlyle Sartor Resartus and Roth American Pastoral

36 of 41

What’s Next?

  • Collaborating with the Social Psychology department to improve our tags
  • Finding shapes of suspense:
    • Genre
    • Historical Period
  • Improving the model:
    • Identifying and analyzing dialogue

37 of 41

Appendix: Extra Material

38 of 41

Tagged Short Stories

  • The Outsider H.P. Lovecraft (1926)
  • The Killers Ernest Hemingway (1927)
  • Three O’Clock Cornell Woolrich (1938)
  • The Quantum of Solace Ian Fleming (1959)
  • The Screwfly Solution James Tiptree, Jr. (1978)
  • The Hook Alvin Schwartz (1981)
  • The Falls George Saunders (1996)

  • Raymond, A Fragment Juvenis (1799)
  • The Anaconda Matthew Lewis (1808)
  • Young Goodman Brown Nathaniel Hawthorne (1835)
  • The Pit and the Pendulum Edgar Allan Poe (1842)
  • The Inhabitant of Carcosa Ambrose Bierce (1887)
  • The Adventure of the Speckled Band Sir Arthur Conan Doyle (1892)
  • The Jolly Corner Henry James (1908)

39 of 41

Expansion

  • Tagging more texts, including novels
  • Refining tagging system
  • Plotting MDWs from short stories across novels - is it possible to generate a map of suspense?

40 of 41

Abigail’s tags

41 of 41

Chelsea’s tags