Variation and Change in Online Social...

Variation and Change in OnlineSocial Networks

Jacob Eisenstein@jacobeisenstein

Georgia Institute of Technology

April 21, 2017

The important of place in online writingyinz ard (“alright”)

lbvs (“laughing but very serious”) ctfu (“cracking the fuck up”)

(Eisenstein et al., 2010, 2014)

DC’s contribution: lls

2009 2010

2011 2012

Language change in the network

How does these macro-scale patternsof variation and change ground out inindividual social networks?

I Can macro-scale phenomena beexplained in terms of individualsocial choices?

I How does network structureaffect the trajectory of languagechange?

Hypotheses

Goel et al. (2016): use large-scale Twitter data to testthree hypotheses about language change.

I H1: language change is transmitted acrosssocial media networks.

I H2: strong ties are better conduits of languagechange.

I H3: geographically local social network ties arebetter conduits of language change (covertprestige; Trudgill, 1972; Bourdieu, 1984).

Dataset

I Twitter analysis is usually conducted on asample from the streaming API (e.g., Eisenstein

et al., 2010; Huang et al., 2016).

I Modeling the fine structure of language changerequires complete data, because randomsamples miss most of the co-occurrences thatreveal sociolinguistic influence.

I This work: a dataset of all public tweetsbetween 2011 and 2014, with 4.35 millionunique user accounts.

Cities and distinctive features

I Atlanta: ain, dese, yeen

I Baltimore: ard, inna, lls, phony

I Charlotte: cookout

I Chicago: asl, mfs

I Los Angeles: graffiti, tfti

I Philadelphia: ard, ctfuu, jawn

I San Francisco: hella

I Washington, DC: inna, lls, stamp

Social network

Two users are considered to have a social networktie if they have each mentioned each other in amessage, e.g.

I User1: @user2 salut

I User2: @user1 what’s up?

This mention network is more socially meaningfulthan the articulated network of follower-followeelinks (Huberman et al., 2008).

Social network

100 101 102 103 104 105

degree

100101102103104105106107

coun

t

mutualoutin

The symmetrized (“mutual”) mention networkyields a more credible degree distribution.

Summary of data

Social network

Bart Lisa

Bart Milhouse

Lisa Homer

Homer Barney

... ...

Locations

Bart Los Angeles

Milhouse Los Angeles

Lisa Atlanta

Homer Chicago

... ...

Language

Bart jawn Feb 1, 2013,

13:45

Milhouse jawn Feb 1, 2013,

13:50

Homer hella Feb 1, 2013,

18:15

Bart lls Feb 2, 2013,

07:30

Milhouse lls Feb 2, 2013,

07:40

... ... ...

Hypotheses

I H1: Language change is transmitted acrosssocial media networks.

I H2: Strong ties are better conduits oflanguage change.

I H3: Geographically local social network tiesare better conduits of language change.

Hypotheses




Shuffle test for influence

t = 1 t = 2

t = 3

I Observed data

P(infection | 0 exposures) =1

4

P(infection |≥ 1 exposures) =2

3

I Randomized data


4

,1

4, . . .


2

,2

3, . . .


t = 2 t = 3

t = 1

I Observed data


4


3

I Randomized data


4

,1

4, . . .


2

,2

3, . . .


t = 3 t = 1

t = 2

I Observed data


4


3

I Randomized data


4,

1

4, . . .


2,

2

3, . . .

Diffusion of innovations

1 2 3+exposures

1.0

1.2

1.4

1.6

1.8

2.0re

lativ

e ris

k

lexicalabbreviationphonetic

I Relative risk: likelihood of infection given exposure,divided by likelihood under random shuffling

I Rel. risk > 1: evidence of non-random contagion.

I For phonetic variables, risk increases with multipleexposures, a characteristic of complex contagion.

Hypotheses




Hypotheses




The Poisson process

I Suppose we have a cascade of event times,{tn}n∈1...N .

I Let y(t1, t2) be the count of events betweentimes t1 and t2. Then,

y(t1, t2) ∼Poisson(Λ(t1, t2)) (1)

Λ(t1, t2) =

∫ t2

t1

λ(t)dt. (2)

The Poisson process

For example:

I y(t1, t2) is the count of the word lls between2013 and 2014

I λ(t) is the (continuously varying) intensityfunction.

Hawkes process

A Poisson process in which the intensity functiondepends on the history (Hawkes, 1971)

λ(t) =µ + α∑tn<t

κ(t − tn), (3)

where,I the time kernel κ decays exponentially witht − tn;

I µ is the base rate;

I α captures the degree of self-excitation.

Hawkes process

For example:

I y(t1, y2) is the count of the word lls

I α captures the tendency of usages of lls to“excite” other usages.

Multivariate Hawkes process

Now suppose each event has some source m.

I The cascade is {(tn,mn)}n∈1...N .

I The intensity for source m is,

λm(t) = µm +∑tn<t

αmn→mκ(t − tn), (4)

where αmn→m is the excitation exerted byevents with source mn on source m.

Multivariate Hawkes process

For example:

I Each source mcorresponds to anindividual social mediauser.

I ym(t1, t2) is the count ofusages of lls by user mbetween t1 and t2.

I αm1→m2 captures theinfluence of m1 on m2.

Parametric Hawkes process

The infection parameters are a linear function ofshared features of each pair of individuals,

αm1→m2= θ>f (m1 → m2). (5)

I We now need estimate only #|θ| parameters,rather than M2.

I We can use features to test hypotheses aboutwhat types of dyads are influential.

Features

F1: Self-excitation δ(m1 = m2)

F2: Social network (m1,m2) ∈ E

F3: Locality (m1,m2) ∈ E andloc(m1) = loc(m2)

F4: Tie strength (m1,m2) ∈ Eand (m1,m2) is astrong tie

Measuring tie strength

Mutual friends

mf (i , j) = #|{k : k ∈ Γ(i) ∩ Γ(j)}| (6)

Adamic & Adar (2003): reweight each mutual friend byits log degree:

aa(i , j) =∑

k∈Γ(i)∩Γ(j)

1

log #|Γ(k)|(7)

We set f4(m1,m2) = 1 if aa(m1,m2) is in the 90thpercentile.

Measuring tie strength

Mutual friends

mf (i , j) = #|{k : k ∈ Γ(i) ∩ Γ(j)}| (6)

Adamic & Adar (2003): reweight each mutual friend byits log degree:

aa(i , j) =∑

k∈Γ(i)∩Γ(j)

1

log #|Γ(k)|(7)

We set f4(m1,m2) = 1 if aa(m1,m2) is in the 90thpercentile.

Hypothesis testing

We compare a series of nested models.

I F2 + F1 vs F1: is language changetransmitted across the social network?

I All features vs F1 + F2 + F4: are local tiesbetter conduits of language change?

I All features vs F1 + F2 + F3: are strong tiesbetter conduits of language change?

Each comparison is performed using a likelihoodratio test, with correction for multiplecomparisons (Benjamini & Hochberg, 1995).

Goodness of fit

103 104 105

Word counts

0

5

10

15

20

25

log

likel

ihoo

d im

prov

emen

t

cookout

graffiti

hella

jawn

phonystamp

ctfuu

llsmfs

tfti

ain

ard

asl

deseinnayeen

Tie strength

lexical

abbreviation

phonetic

103 104 105

Word counts

cookoutgraffitihella

jawn

phonystampctfuu

lls

mfstfti

ain ard

asl

deseinnayeen

Local

I Tie strength feature improves model fit for many words,evidence for H2.

I Geography locality rarely improves model fit,contradicting H3.

Why are social media variables so local?

No support for the hypothesisthat exposures from local tiesare especially influential.

But! Most exposures are local. Two reasons:

1. Most ties are local

2. Audience design (Pavalanathan & Eisenstein, 2015)

Why are social media variables so local?

No support for the hypothesisthat exposures from local tiesare especially influential.

But! Most exposures are local. Two reasons:

1. Most ties are local

2. Audience design (Pavalanathan & Eisenstein, 2015)

Broadcast

Hashtag-initial

Addressed

Logistic regression

I Dependent variable: does the tweet containa non-standard, geographically-specific word(e.g., lbvs, hella, jawn)

I PredictorsI Message type: broadcast, addressed, #-initialI Controls: message length, author statistics

Small audience → less standard language

Distinguishing local ties

To distinguish local audiences:

I Use GPS metadata to identify author locationsI Associate metro m with user u if u is

@-mentioned by:I at least three users within metro m;I nobody outside metro m.

The social network lets us impute the locations ofunknown users from the 1-2% of users who revealtheir GPS! (Sadilek et al., 2012)

Distinguishing local ties

To distinguish local audiences:

I Use GPS metadata to identify author locationsI Associate metro m with user u if u is

@-mentioned by:I at least three users within metro m;I nobody outside metro m.

The social network lets us impute the locations ofunknown users from the 1-2% of users who revealtheir GPS! (Sadilek et al., 2012)

Local audience → less standard language

Local ties make non-standard language even morelikely.

Local audience → less standard languageM

ore

men

tion

s by

use

rs in

same

met

ro a

rea

More mentions by users in other metro areas

Messages containing local variable

Messages not containing local variable

Next steps

I How to find geographically distinctive linguisticvariables? (Nguyen & Eisenstein, 2017)

I What is the effect of incomplete data onestimates of linguistic influence?

I What is the impact of technological mediationon online language?

Next steps

I How to find geographically distinctive linguisticvariables? (Nguyen & Eisenstein, 2017)

I What is the effect of incomplete data onestimates of linguistic influence?

I What is the impact of technologicalmediation on online language?

Emojis vs emoticons

Pavalanathan & Eisenstein (2016): How has theintroduction of emojis impacted the use ofnon-standard language?

Causal inference design

I Treatment: adoption of emojis

I Outcome: frequency of emoticons and othernon-standard tokens

Emojis vs emoticons

After treatment, emojiadopters use:

I significantly feweremoticons(0.40%→ 0.12%)

I significantly morestandard tokens(85.2%→ 86.2%)

Which emoticons?

Nose :-) :-/Horizontal O O - -Equal sign eyes =) =DWink eyes ;-) ;DTongue :P ;PSlant mouth :/ :-Smiling mouth :) ;)Laughing mouth :D ;-D’O’ mouth :O :-OTears :’) :’(Reversed (-: (:

Which emoticons?

Nose :-) :-/Horizontal O O - -Equal sign eyes =) =DWink eyes ;-) ;DTongue :P ;PSlant mouth :/ :-Smiling mouth :) ;)Laughing mouth :D ;-D’O’ mouth :O :-OTears :’) :’(Reversed (-: (:

Language variation: a challenge for NLP

“I would like to believe he’ssick rather than just meanand evil.”

“You could’ve been gettingdown to this sick beat.”

(Yang & Eisenstein, 2017)

Language variation: a challenge for NLP

“I would like to believe he’ssick rather than just meanand evil.”

“You could’ve been gettingdown to this sick beat.”

(Yang & Eisenstein, 2017)

Personalization by ensemble

I Goal: personalized conditional likelihood,P(y | x , a), where a is the author.

I Problem: We have labeled examples for only afew authors.

I Personalization ensemble

P(y | x , a) =∑k

Pk(y | x)πa(k)

I Pk(y | x) is a basis modelI πa(·) are the ensemble weights for author a

Personalization by ensemble

I Goal: personalized conditional likelihood,P(y | x , a), where a is the author.

I Problem: We have labeled examples for only afew authors.

I Personalization ensemble

P(y | x , a) =∑k

Pk(y | x)πa(k)

I Pk(y | x) is a basis modelI πa(·) are the ensemble weights for author a

Homophily to the rescue?

Sick! Sick!

Sick!Sick!

Labeled data

Unlabeled data

Are language styles assortative on the socialnetwork?

Evidence for linguistic homophilyPilot study: is classifier accuracy assortative on theTwitter social network?

assort(G ) =1

#|G |∑

(i ,j)∈G

δ(yi = yi)δ(yj = yj)

+ δ(yi 6= yi)δ(yj 6= yj)

0 20 40 60 80 100rewiring epochs

0.7000.7050.7100.7150.7200.7250.7300.735

asso

rtativ

ity

follow


mention


retweet

original networkrandom rewiring

Evidence for linguistic homophilyPilot study: is classifier accuracy assortative on theTwitter social network?

assort(G ) =1

#|G |∑

(i ,j)∈G

δ(yi = yi)δ(yj = yj)

+ δ(yi 6= yi)δ(yj 6= yj)


0.7000.7050.7100.7150.7200.7250.7300.735

asso

rtativ

ity

follow


mention


retweet

original networkrandom rewiring

Network-driven personalization

I For each author, estimatea node embeddingea (Tang et al., 2015).

I Nodes who shareneighbors get similarembeddings.

πa =SoftMax(f (ea))

P(y | x , a) =K∑

k=1

Pk(y | x)πa(k)

Results

Mixture of Experts NLSE Social Personalization

0.0

0.5

1.0

1.5

2.0

2.5

3.0F1

impr

ovem

ent o

ver C

onvN

et

+0.10

+1.90

+2.80

Twitter Sentiment Analysis

Improvements over ConvNet baseline:

I +2.8% on Twitter Sentiment Analysis

I +2.7% on Ciao Product Reviews

NLSE is prior state-of-the-art (Astudillo et al., 2015).

Variable sentiment words

More positive More negative

1 banging loss fever brokenfucking

dear like god yeah wow

2 chilling cold ill sick suck satisfy trust wealth stronglmao

3 ass damn piss bitch shit talent honestly voting winclever

4 insane bawling fever weird cry lmao super lol haha hahaha

5 ruin silly bad boring dreadful lovatics wish beliebers ariana-tors kendall

Conclusions

I Internet media continues to create new socialconfigurations, new subcultures, and newcommunicative affordances.

I New opportunities for studying themicro-foundations of language change.

I A future of blurred lines: online vs IRL, humancontrol vs autonomy, text vs speech.

Acknowledgments

I Collaborators: Ming-Wei Chang, FernandoDiaz, Naman Goyal, Rahul Goel, VinodhKrishnan, John Paparrizos, UmashanthiPavalanathan, Sandeep Soni, Hanna Wallach,Yi Yang

I Sponsors: National Science Foundation, AirForce Office of Scientific Research, NationalInstitutes for Health, Microsoft Research.

References IAdamic, L. A. & Adar, E. (2003). Friends and neighbors on the web. Social networks, 25(3), 211–230.

Astudillo, R. F., Amir, S., Lin, W., Silva, M., & Trancoso, I. (2015). Learning word representations from scarce andnoisy data with embedding sub-spaces. In Proceedings of the Association for Computational Linguistics(ACL), Beijing.

Benjamini, Y. & Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach tomultiple testing. Journal of the Royal Statistical Society. Series B (Methodological), 289–300.

Bourdieu, P. (1984). Distinction: A social critique of the judgement of taste. Harvard University Press.

Eisenstein, J., O’Connor, B., Smith, N. A., & Xing, E. P. (2010). A latent variable model for geographic lexicalvariation. In Proceedings of Empirical Methods for Natural Language Processing (EMNLP), (pp. 1277–1287).

Eisenstein, J., O’Connor, B., Smith, N. A., & Xing, E. P. (2014). Diffusion of lexical change in social media. PLoSONE, 9.

Goel, R., Soni, S., Goyal, N., Paparrizos, J., Wallach, H., Diaz, F., & Eisenstein, J. (2016). The social dynamics oflanguage change in online networks. In The International Conference on Social Informatics (SocInfo).

Hawkes, A. G. (1971). Spectra of some self-exciting and mutually exciting point processes. Biometrika, 58(1),83–90.

Huang, Y., Guo, D., Kasakoff, A., & Grieve, J. (2016). Understanding us regional linguistic variation with twitterdata analysis. Computers, Environment and Urban Systems, 59, 244–255.

Huberman, B., Romero, D. M., & Wu, F. (2008). Social networks that matter: Twitter under the microscope.First Monday, 14(1).

Nguyen, D. & Eisenstein, J. (2017). A kernel independence test for geographical language variation.Computational Linguistics, in press.

Pavalanathan, U. & Eisenstein, J. (2015). Audience-modulated variation in online social media. American Speech,90(2).

Pavalanathan, U. & Eisenstein, J. (2016). More emojis, less :) The competition for paralinguistic functions inmicroblog writing. First Monday, 22(11).

References II

Sadilek, A., Kautz, H., & Bigham, J. P. (2012). Finding your friends and following them to where you are. InProceedings of the Conference on Web Search and Data Mining (WSDM), (pp. 723–732).

Tang, J., Qu, M., Wang, M., Zhang, M., Yan, J., & Mei, Q. (2015). Line: Large-scale information networkembedding. In Proceedings of the Conference on World-Wide Web (WWW), (pp. 1067–1077).

Trudgill, P. (1972). Sex, covert prestige and linguistic change in the urban british english of norwich. Language inSociety, 1(2), 179–195.

Yang, Y. & Eisenstein, J. (2017). Overcoming language variation in sentiment analysis with social attention.Transactions of the Association for Computational Linguistics (TACL).

Date post:	19-May-2018
Category:	Documents
Upload:	vuongkhuong
View:	216 times
Download:	2 times

Variation and Change in Online Social...

Documents