/
Style and Influence in Social Text Style and Influence in Social Text

Style and Influence in Social Text - PowerPoint Presentation

debby-jeon
debby-jeon . @debby-jeon
Follow
375 views
Uploaded On 2018-01-18

Style and Influence in Social Text - PPT Presentation

112729 Announcement Project reports next week same drill as midterm reports reverse order as midterm reports W e know youre not done yet but you will be by midnight Mon 1210 right ID: 624687

people words frequent admins words people admins frequent style text status word literary signals common summary topics today

Share:

Link:

Embed:

Download Presentation from below link

Download Presentation The PPT/PDF document "Style and Influence in Social Text" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.


Presentation Transcript

Slide1

Style and Influence in Social Text

11-27-29Slide2

Announcement

Project reports next week

same drill as midterm reports

reverse order as midterm reports

W

e know you’re not done yet

… but you will be by midnight Mon 12/10, right?

start with one slide summarizing midtermSlide3

FCE’s

Are now open

We do read them…and people do care

Especially this year

free-text comments on assignments/structure/layout of course

very

welcomeSlide4

Puzzle time

Ths

sntnc

hs

n

vwls

i

eee

a o

osoaSlide5

Today’s topics

Summary:

there are

signals

in

common words

What can you

infer

from how

people use the most frequent words in text?Slide6

Today’s topics

Summary

: there are

signals

in

common words

What can you

infer

from how

people use the most frequent words

in

text

?Slide7

Today’s topics

Summary: there are signals in common words

What can you infer from how people use the most frequent words in text?Slide8

Today’s topics

Summary: there are signals in common words

What can you infer from how people use the most frequent words in text?

Patterns of usage

 ”literary style”

predicts: authorship, gender, …

Style changes according to situation

and is transmitted from person to person

Outline:

some background and two recent papersSlide9
Slide10

Background: Authorship attribution

Mosteller

and Wallace, 1964. “Inference and Disputed Authorship”: frequency of function words can be used to classify documents by author.

Function words are not under conscious control

Function word use is independent of content

Histogram of function words is okSlide11

Authorship attribution

Schlomo

Argamon

,

Schlomo

Levitan

SVM on histogram

of 200 most frequent wordsSlide12

COLING 2006Slide13
Slide14

LIWC

1986: writing about emotional upheavals improved physical health (!)

Can you refine this statement?

what

sort

of writings yield the

best

results?

but: people don’t agree on ratings

and: “judges tend to get depressed when reading depressing stories.”so: design an automatic “instrument” to rate writings (Linguistic Inquiry and Word Count) based on most frequent wordsSlide15

LIWC words - cover about 55% of the tokens (not types) in most text

Categories are mostly designed by hand, by committeeSlide16
Slide17
Slide18
Slide19
Slide20

Another signal of rank: starting a fashionSlide21

most frequent 200 words

Is literary style like fashion? Can you track literary influence? Can you find high-status, influential people by modeling literary style?Slide22

most frequent 200 wordsSlide23
Slide24

People adopt each other’s mannerisms and style in many ways….Slide25
Slide26
Slide27

Corpus

Pennebaker

&

Niederhoffer

, 2002:

98 pairs in the lab + Watergate tapes

Twitter A:

1.3M “conversations” between 300k users--many are too short to analyze successfully

Twitter B: More crawling

all pairs with 2+ conversationsall posts from these pairs15M tweets, 7800 users, 215k conversations, 2200 pairsSlide28

Measuring “cohesion” for a property CSlide29

Measuring “cohesion”

Tweet T contains word from class C

Reply R contains word from class C

T and R are a “turn”Slide30
Slide31

Measuring “accommodation” and “influence”

T

b

, from

b,

is a reply to

T

a

, from

a Slide32

T

b

uses word class C in a reply to

a

T

b

uses word class C in a reply to

a

after

a

uses CSlide33

Evidence of

fashion

in linguistic style spreading through a conversation

Time lag suggests

influence

not

associative sorting

We don’t have anything like direction…..Slide34

If

Acc

(

a,b

)>0:

Symmetric:

Acc

(

b,a

) > 0

Default asymmetric:

Acc

(

b,a

) = 0

Divergent asymmetric:

Acc

(

b,a

) < 0Slide35

Does one party accommodate more than the other?

Accommodation does

not

correlate with “status” features like #followers, #days on Twitter, ….Slide36

????

Does one party accommodate more than the other?Slide37
Slide38

Datasets

Wikipedia:

wikipedia

editors

talk

pages: 240k conversations; plus 32k discussions over who gets promoted to admins.

Status: admin

vs

non-admin

Dependence: learning to support/rejectSupreme court: 50k verbal exchanges for 204 cases.Status: chief justice vs justice vs lawyerDependence

: leaning to support/learning to rejectSlide39

Experiments

Similar notion of “coordination” (=

accomodation

)

Hypotheses:

e.g., you accommodate more when speaking to a big shot

and he coordinates less with other peopleSlide40
Slide41

more coordination with admins than non-admins

admins coordinate

more

with others than non-adminsSlide42

admins coordinate

more

with others than non-admins

Why?

Maybe the folks that become admins are different somehow?

eg

more accommodating?Slide43

the people that

eventually become admins

coordinate more than people

who

eventually fail to become adminsSlide44

revised hypothesis:

after

you become an admin you will coordinate with others

less

than you did beforeSlide45

What about the court dataset?Slide46

What about the court dataset?Slide47

Status prediction

Given conversation between

x,y

predict if

status(x)>status(y)

or vice-versa

Very easy to do in Supreme Court domain (“your honor,….”)

Hard for humans in Wikipedia (inter-annotator

aggrement ~= 80%, accuracy ~=70%)Slide48
Slide49

One more observation…Slide50
Slide51

So to summarize…

Summary: there are signals in common words

Even though we don’t think about how we use them

Patterns of usage

 ”literary style”

predicts: authorship, gender, …

Style changes according to situation

and is transmitted from person to person

you can observe that transmission (accommodation, coordination) and determine its direction

the direction of accommodation it tells you something about the status of the speakers