A Long-Term Analysis of Polarization on Twitter

A Long-Term Analysis of Polarization on Twitter Kiran Garimella Ingmar Weber

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

A Long-Term Analysis of Polarization on Twitter 2017 ICWSM / arXiv 1703.02769 garimella2017-polarization garimella2017-polarization.txt, the project's staged copy ef47b538ceff7de5173304334d67288616c18b6b arxiv-open political-science

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them existence Quotation withheld because this record is not cleared for redistribution. political polarization on Twitter Quotation withheld because this record is not cleared for redistribution. observational-study observational-study Quotation withheld because this record is not cleared for redistribution. mechanisms the time a user began following is ESTIMATED, from the reverse-chronological order the API returns followers in, lower-bounded by each follower's account creation date Quotation withheld because this record is not cleared for redistribution. the paper's REASON for the Bayesian treatment - that a single act from one side is a weak signal - is the sentence immediately before this one, and widening to reach it pulls in a footnote the extraction has interleaved into the body ('4Since the Twitter API restricts us to the last 3200 tweets'), so the tag is trimmed instead each user's leaning is a beta-binomial coin-bias estimate - a uniform prior over the two sides, then one increment per follow or retweet Quotation withheld because this record is not cleared for redistribution. mechanisms polarization is defined as DISTANCE FROM BALANCE rather than as leaning, so a strongly left and a strongly right user score identically Quotation withheld because this record is not cleared for redistribution. mechanisms a hashtag inherits its leaning from the users who used it that week, and users who did not retweet a seed account that week contribute nothing Quotation withheld because this record is not cleared for redistribution. mechanisms two explicit noise controls chosen because the early years are thin: a minimum of five users per hashtag, and a five-week moving average Quotation withheld because this record is not cleared for redistribution. mechanisms automated and bad-faith accounts are handled BY THE PASSAGE OF TIME rather than by detection - the study relies on Twitter having suspended them in the years before collection Quotation withheld because this record is not cleared for redistribution. Twitter API 'most recent activity only' limitations hinder longitudinal study what this limits: any longitudinal claim: API LIMITATIONS HINDER SUCH STUDIES - only most-recent activity is retrievable Quotation withheld because this record is not cleared for redistribution. Automated tweets and astroturfing were not detected or removed. what this limits: the authenticity of the measured activity: WE DID NOT EXPLICITLY ATTEMPT TO DETECT ASTROTURFING AND OTHER TYPES OF AUTOMATED TWEETS Quotation withheld because this record is not cleared for redistribution. Hashtag hijacking could still contaminate the content measure. what this limits: the content measure, with the boundary decision stated: HASHTAG HIJACKING COULD STILL AFFECT OUR ANALYSIS, THOUGH THESE EFFECTS ARE ARGUABLY ALSO PART OF THE POLITICAL LANDSCAPE AND SHOULD BE INCLUDED Quotation withheld because this record is not cleared for redistribution. The study cannot settle whether social media causes, reflects or reduces offline polarization. what this limits: any causal reading: the study cannot settle whether social media causes, reflects or reduces offline polarization Quotation withheld because this record is not cleared for redistribution. The election-period pattern in hashtag polarization does not hold for the 2010 midterms, probably because of noisier early estimates. what this limits: the election-period pattern: FOR THE 2010 MIDTERM ELECTIONS, HOWEVER, THIS OBSERVATION DOES NOT HOLD, POTENTIALLY DUE TO NOISIER ESTIMATES BASED ON FEWER ACTIVE USERS Quotation withheld because this record is not cleared for redistribution. Pre-September-2009 tweets were excluded because their volume was too low. what this limits: the temporal coverage of the analysis: WE ONLY CONSIDER TWEETS FROM SEPTEMBER 2009 ONWARDS ... SINCE THE VOLUME FOR EARLIER TWEETS IS LOW Quotation withheld because this record is not cleared for redistribution. The follow times underlying the whole time series are estimated rather than observed, and the accuracy attributed to that estimator is reported second-hand and only for accounts with millions of followers. what this limits: the year-by-year following curve: follow dates are estimated by the method of Meeder et al., whose accuracy is reported to hold for users with millions of followers, while the analysis uses all cases with estimated follow dates from January 2009 onwards Quotation withheld because this record is not cleared for redistribution. The content measure is not specified in this paper; its method is applied from another work and the reader is sent there for the details. what this limits: the hashtag-polarization series: the procedure that produces it - leaning assignment, volume correction and smoothing - is taken from another paper, and this one refers the reader there for details Quotation withheld because this record is not cleared for redistribution. Each user's polarization value is a posterior belief that begins from an assumed balance between the two sides, not a summary of that user's observed behaviour alone. what this limits: a per-user polarization value: it is the mean of a beta posterior that starts from a uniform prior over left and right and is updated by each observed follow or retweet, so a user with few observed actions is scored largely by the prior Quotation withheld because this record is not cleared for redistribution. social-network-structure analysis applied to measure political polarization on Twitter Quotation withheld because this record is not cleared for redistribution. three separate questions, each answered from A DIFFERENT TYPE OF INFORMATION, rather than one measure carried through eliminates: a single data source's idiosyncrasies standing in for the phenomenon - polarisation measured only through, say, retweets is a fact about retweeting until a second and third instrument agree · outcome (stated by the tagger, no quotation): held Quotation withheld because this record is not cleared for redistribution. echo chamber echo chamber Quotation withheld because this record is not cleared for redistribution. polarization polarization Quotation withheld because this record is not cleared for redistribution. leaning leaning Quotation withheld because this record is not cleared for redistribution. hashtag polarization hashtag polarization Quotation withheld because this record is not cleared for redistribution. polarization polarization Quotation withheld because this record is not cleared for redistribution. polarization - STIPULATED AS RESTRICTION, not as hostility or extremity: a tendency to be restricted to one side of the spectrum when obtaining or engaging with political information Quotation withheld because this record is not cleared for redistribution. polarization - AS A NUMBER: twice the absolute deviation of a beta-posterior leaning from 0.5, so it measures distance from a balanced diet on a 0.0 to 1.0 scale and nothing else Quotation withheld because this record is not cleared for redistribution. polarization - AS LEXICAL DIVERGENCE: two sides that describe the same thing in completely different words, the paper's own example being global warming versus climategate Quotation withheld because this record is not cleared for redistribution. the term's scope - FIXED BY STIPULATION IN A FOOTNOTE: every later use of the bare word is declared to mean political polarization Quotation withheld because this record is not cleared for redistribution. builds-on → the hashtag-leaning methodology of [15] Quotation withheld because this record is not cleared for redistribution. builds-on → Meeder et al. [10] follow-time estimation method Quotation withheld because this record is not cleared for redistribution. extends → Andris et al. [2] (partisanship of the US Congress) Quotation withheld because this record is not cleared for redistribution. extends → Liu et al. [9] (seven-year longitudinal Twitter analysis) Quotation withheld because this record is not cleared for redistribution. polarization —operationalised-as→ leaning polarization leaning strength of the relation: 1 Quotation withheld because this record is not cleared for redistribution. hashtag polarization —is-a→ polarization hashtag polarization polarization strength of the relation: 1 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but endpoint by anaphora: the section is headed 'Hashtag Polarization'; the span writes 'polarization of hashtags'. Quotation withheld because this record is not cleared for redistribution. 679,000 — size of the longitudinal Twitter dataset Quotation withheld because this record is not cleared for redistribution. 10%-20% — relative increase in online polarization over eight years, depending on measure Quotation withheld because this record is not cleared for redistribution. eight years — span of the longitudinal analysis (two presidential + two mid-term elections) about this quote: QUOTE REPAIRED: the extractor split a hyphen-space and the quote was recorded faithfully with the damage in it, which made it unanchorable. The repaired string is an EXACT SUBSTRING of the staged text - that is why it was accepted, not because a rule produced it. The text as originally recorded is kept in evidence as extracted. Quotation withheld because this record is not cleared for redistribution. > 60% — share of Americans who get news from social media (Pew) Quotation withheld because this record is not cleared for redistribution. Online political polarization increased over the past eight years, a relative change of 10-20% depending on the measure. Quotation withheld because this record is not cleared for redistribution. echo chamber echo chamber Quotation withheld because this record is not cleared for redistribution. Bayesian Bayesian about this quote: QUOTE REPAIRED: the extractor split a hyphen-space and the quote was recorded faithfully with the damage in it, which made it unanchorable. The repaired string is an EXACT SUBSTRING of the staged text - that is why it was accepted, not because a rule produced it. The text as originally recorded is kept in evidence as extracted. THE QUOTE CONTINUES A SENTENCE THAT THIS COPY CUTS IN TWO, AND ITS FIRST WORDS ARE NOT PART OF THAT SENTENCE - a footnote stands between it and the words it follows, which read 'not necessarily a strong signal', and the quote opens with the end of that footnote's split URL. Quotation withheld because this record is not cleared for redistribution. mean mean about this quote: QUOTE REPAIRED: the extractor split a hyphen-space and the quote was recorded faithfully with the damage in it, which made it unanchorable. The repaired string is an EXACT SUBSTRING of the staged text - that is why it was accepted, not because a rule produced it. The text as originally recorded is kept in evidence as extracted. Quotation withheld because this record is not cleared for redistribution. p-value p-value Quotation withheld because this record is not cleared for redistribution. t-test t-test about this quote: QUOTE REPAIRED: the extractor split a hyphen-space and the quote was recorded faithfully with the damage in it, which made it unanchorable. The repaired string is an EXACT SUBSTRING of the staged text - that is why it was accepted, not because a rule produced it. The text as originally recorded is kept in evidence as extracted. Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://arxiv.org/abs/1703.02769.