Slide 1

S AY A NYTHING Selma P. and Parth P.

Slide 2

S AY W HAT ? Data-driven system that allows for interactive storytelling Human author takes turns with computer author

Slide 3

E XAMPLE H: Youll never believe what happened last night. C: Leigh laughed at my joke but I couldnt help but think liz would have laughed harder. H: The joke wasnt very funny in a ha ha kind of way C: It wasnt anything like that, I thought he was going to give me a good night kiss but he ended up licking my cheek, she declared. H: It made me sneeze and snort out loud. C: And now my nose hurts form the snorting.

Slide 4

E XAMPLE H: Youll never believe what happened last night. C: Leigh laughed at my joke but I couldnt help but think liz would have laughed harder. H: The joke wasnt very funny in a ha ha kind of way C: It wasnt anything like that, I thought he was going to give me a good night kiss but he ended up licking my cheek, she declared. H: It made me sneeze and snort out loud. C: And now my nose hurts form the snorting.

Slide 5

E XAMPLE H: Youll never believe what happened last night. C: Leigh laughed at my joke but I couldnt help but think liz would have laughed harder. H: The joke wasnt very funny in a ha ha kind of way C: It wasnt anything like that, I thought he was going to give me a good night kiss but he ended up licking my cheek, she declared. H: It made me sneeze and snort out loud. C: And now my nose hurts form the snorting.

Slide 6

W HY DO THIS ? Intriguing Blends structure with creativity Void Game of sorts Fits in the space between language games and more graphically oriented video games Foundational work in progress Research

Slide 7

N ARRATIVE M ODELS (M AJEWSKI 2003) Five ways: Linear String of pearls Branching Amusement park Building blocks

Slide 8

W HAT S BEEN DONE BEFORE ? Faade http://interactivestoriesonline.com/ http://www.interactivenarratives.org/ Top Down Approach Strict domains Technical/ non-inclusive

Slide 9

H OW IT WORKS 1. User enters sentence 2. Users sentence is used to search corpus using term frequency - inverse document frequency (tf- idf) algorithm 3. Highest scored match is retrieved. Sentence after best match is outputted by computer 4. Repeat

Slide 10

H OW DID THEY DO IT ?

Slide 11

G ETTING D ATA ( DB ) Considered Manual (StoryCorps/ Fed. Writers) Well curated biased Favored blog posts (3.4 million) 1.06 billion words Extraction Only 17% textual material on weblogs is narrational 3.7 million story segments 66.5 million sentences Favored Recall over Precision FN preferred over FP

Slide 12

G ETTING D ATA Spinn3r.com 44 million weblog. Sampled / hand annotated 5,270 blogs. Annotated validations set

Slide 13

G ETTING D ATA Randomized training/ testing Supervised machine learning Binary classification problem Confidence Weighted Linear Classifier [Dredze 2008]

Slide 14

G ETTING D ATA Rando. Data Set annotation training/ testing F(x)F(x) Sampling

Slide 15

C ORPUS ( DB ) C REATION Crawled blogs and applied algorithm: (2012): Post- Processing Parse trees Verb tenses First personal pronouns

Slide 16

C ORPUS ( DB ) C REATION Rando. Data Set crawl blogs annotation training/ testing F(x)F(x) Sampling Blog data F(x)F(x) Story data

Slide 17

A PPLICATION Querying Corpus ( ) Optimized: Return a list of stories that contain any matching words with user input Use TF-IDF ! Story data

Slide 18

A PPLICATION Rando. Data Set crawl blogs annotation training/ testing F(x)F(x) Sampling Blog data F(x)F(x) Story data User Input Query Story data Match Algorithm (tf id) Computer Output

Slide 19

T ERM FREQUENCY INVERSE DOCUMENT FREQUENCY ( TF - IDF ) TF-IDF is a numerical statistic that reflects how important a word is to a document in a corpus The term frequency measures how often a word appears in a document The inverse document frequency is a measure of how common a word is within the corpus as a whole. It tells us how much information a word provides.

Slide 20

T ERM FREQUENCY INVERSE DOCUMENT F REQUENCY ( TF - IDF ) Image credit: Li(2011)

Slide 21

R ESULTS

Slide 22

A REAS TO I MPROVE Metrics? Entertainment Coherence

Slide 23

A REAS TO I MPROVE Metrics? Entertainment Coherence Believability / Usability Compare how well with next sentence user would have written

Slide 24

A REAS TO I MPROVE Fail to use all preceding sentences. Only returns highest ranked search.

Slide 25

A REAS TO I MPROVE Fail to use all preceding sentences. Only returns highest ranked search.

Slide 26

A REAS TO I MPROVE Fail to use all preceding sentences. Only returns highest ranked search.

Slide 27

F UTURE WORK No narrative plot

Slide 28

W ORK IN P ROGRESS Foundational Last article was 2012

Slide 29

T HANKS

Slide 30

D ISCUSSION Q S

Slide 31

I MPROVEMENTS ?

Slide 32

Q UALITY A SSESSMENT

Slide 33

M ODIFY C ORPUS

Slide 34

P APER C ONTENT ?