Introduction

About Us

:::: {.columns .top}

::: {.col}

Wellesley College

Wellesley, Massachusetts, U.S.A.

• Sarah Yan (syan@wellesley.edu)
  Graduated Spring 2019
• Peter Mawhorter (pmawhort@wellesley.edu)
  Computer Science Lab Instructor

:::

::: {.col} {style=“border-radius: 8pt; width: 370px; max-width: 1000px; margin-top: -12pt; margin-left: 6pt;”}  :::

::::

Background

Esports

• Competitive multiplayer online games.
• Team-based competition, organized like a sport.
• Millions of viewers globally for playoff events.
• Smaller than traditional sports, but growing quickly.
• Younger audiences that engage on social media.

::: {.attribution} Sources: https://www.hotspawn.com/guides/esports-sports-how-the-two-compare/,
https://www.slideshare.net/ActivateInc/activate-tech-media-outlook-2018 (slides 87 and 90),
What is eSports and why do people watch it? (Hamari & Sjöblom 2017) :::

Overwatch

• Released in 2016 by Blizzard Entertainment
• Tens of millions of players.
• Hundreds of thousands of viewers of the Overwatch League (OWL).
• 20 teams loosely associated with cities in North America, Europe, and East Asia.
• League play has moved online during the pandemic.

::: {.attribution} Sources: Overwatch League Announcement and Overwatch League 2019 Fact Sheet from https://blizzard.gamespress.com/en/Esports-Overwatch#?tab=documents-7 :::

Fans & Social Media

:::: {.columns .top}

::: {.col} - An important site of fan identity formation. - Potential for measurement of audience reception. - Site of marketing & team outreach. :::

::::

::: {.attribution} E.g.: Sport Fandom in the Digital World (Pegoraro 2012) & From Loving the Hero to Despising the Villain: Sports Fans, Facebook, and Social Identity Threats (Sanderson 2013) :::

Sentiment Analysis

• Text analysis to extract opinions.
• Often applied to Twitter data.
• Open-source libraries like the Natural Language ToolKit and TextBlob
• Focus on positive vs. negative opinion of a subject.

::: {.attribution} Sources: Opinion Mining and Sentiment Analysis (Pang & Lee 2008) Sentient Analysis of Twitter Data (Agarwal et al. 2011) :::

Research Questions

Big Questions

• What can social media tell us about synchronous fan engagement with an (e)sports event?
  • How does team performance affect engagement?
  • Can pre-match opinions predict match outcomes?
  • Can social media reveal fan psychology?
• Can automatic analysis answer these questions?

::: {.attribution} Related work: Real‐Time Event Summarization of Esports Events Through Twitter Streams,
Real-time nonverbal opinion sharing through mobile phones during sports events (Shirazi et al. 2011),
Text Mining of Audience Opinion in eSports Events (Omella Mainieri et al. 2017) :::

Small Question

Are tweets that engage with a particular team on average more positive for the winning team in a match?

::: {.small} (Assumption: more tweets will come from fans of the team than from their adversaries.) :::

::: {.attribution} (Berntsen (chapter 9) in Flashbulb Memories: New Challenges and Future Perspectives, 2017) :::

Methodology Question

Are easy-to-use open-source libraries sufficient for analyzing Twitter sentiment?

::: {.small} (We used TextBlob which is built on top of NLTK along with Tweepy.) :::

Approach

Data Collection

• Twitter API search feature (via Tweepy).
• Looked for team mentions and hash tags.
• Filtered out some terms to exclude unrelated tweets.
• Collected results before, during, and shortly after matches (12 hours following the start of a match).
• Across 25 matches: average of 36 tweets pre-match and 276 in the match period (per team).
• High variance: 0-201 tweets pre-match and 26-754 in the match period.

Analysis

• Extracted tweet text, ignoring emoticons.
• Used TextBlob’s built-in sentiment analysis directly.
  • Part-of-speech tagging plus a lexicon of per-word sentiments.
• Each tweet gets a polarity score between -1 and +1.

::: {.tiny} Tweet Polarity —— ——— @BostonUprising it was hard fought but I know you will keep improving and surprising everyone. 0.204166667 @Closer @SFShock Nah. Y’all did good https://t.co/YDeoVeGHAp 0.7 @Crimzo @SFShock You guys popped off 0 This game is kind of sad. Numbani sucks to attack on anyway, rip Outlaws. -0.15 @Outlaws Hey outlaws, a tip of advice. Fire that sorry sack of crap you call a coach. Sincerely, a former fan -0.433333333 I BELIEVED. I BELIEVED THEY CAN WIN WAHOOOOOOOO https://t.co/aGChgT4Kex 0.8 watching seoul dynasty is such a rollercoaster of emotions #TigerNation https://t.co/g9MRqJ0pWd 0 :::

Analysis

• Used only match-period data (ignored pre-match).
• Each game has winning and losing sides.
• Average polarities separately across mentions of the winning team and the losing team.
• We expected higher average polarity for the winning side in each match than for the losing side.

Results & Discussion

p = 0.158

[We did not find a significant difference between the average polarity of tweets that engaged with winning teams vs. losing teams.]{.fragment}

Why?

• Our hypothesis could be misguided
  (e.g., flaming the winner).
• Maybe we just need more data?
• Do we need to refine our methods?

What should we expect?

• Who do we expect to express positive sentiment?
• When do we expect positive sentiment?
• What about negative sentiment?
• More nuanced theories of fan engagement would be helpful here.

How much data do we need?

• 250 tweets per match is not a lot.
• Inaccuracies in sentiment tagging or tweet filtering should be susceptible to the law of large numbers.
• As we collected more data, the p-value for our hypothesized relationship did shrink.

Are our results accurate?

• Sentiment is highly contextual
  (e.g., a disappointed fan vs. a taunt from a fan of the opposing team).
• Twitter data is complex (Should we include emoticons? What about sarcasm?)
• Fan communities use idiosyncratic language.

How should we analyze the data?

• Different choices for filtering or data collection
• More complicated sentiment analysis techniques
  • Why haven’t they become the default?

Takeaways & Future Work

Next Steps

• Gather more data.
  • Current infrastructure + other audiences?
  • More details (fan identity?)
• Explore sentiment analysis methods.
• Test more hypotheses.

Takeaways

• Fan engagement is complicated.
  • Demands nuanced data.
  • Understanding smaller communities likely requires a qualitative approach.
• Sentiment analysis is not straightforward.
  • A one-size-fits-all approach might not be adequate.