Crossroads The ACM Magazine for Students

Sign In

Association for Computing Machinery

Magazine: Features Beyond freebird

back to top 

Four years ago, fresh out of graduate school, I started my first job as a research scientist at a large Internet company. People, social networks, media, and what was soon to be known as "big data" was on a slow rise to its boiling point. The Sunday before I started work was the final match of the 2006 World Cup: Italy versus France. Not even unpacked yet, I made my way to Dolores Park in San Francisco where I heard the game would be shown. Some 3,000 people were estimated to attend the event in the medium-sized park. I arrived to the crowd 10,000 strong.

It was so crowded that I could hardly make out Zidane's game-ending red card from my vantage point on top of a hill. After the game, I wondered why so many people would sardine themselves to watch but not really watch the game. What socially drives people to experience media together? Why is sharing conversation in context as important as watching the game itself? How does the conversation change how we perceive the event?

Fast forward to the 2010 World Cup. People gathered worldwide in bars, restaurants, homes, parks, and other places to watch the matches. Something new happened. In real time, people also shouted the cheers, the agony, and the "GOOOOOAAAALLL"s across the internet via Flickr photos, social network updates, and microblog posts. Twitter, microblogging's principal representative, saw record-breaking post numbers just south of 3,000 tweets a second where their normal average is about 750. The mere fact that some 2,000 people a second spread around the world can converge around a single event broadcast from Johannesburg is incredible.

But, how do we start to understand what it means? How does one soccer match project meaning into a largely disconnected sea of voices? Trivially, one could suggest a spike in people saying "goal" means a goal occurred. This would likely work for simple event detection, like identifying a rock concert encore by the swell of people shouting, "Freebird," regardless of what band is playing.

Elections, concerts, and sporting matches all carry online conversations which, when viewed only from a Twitter stream, envelop the event and alter our perception. We should aim to understand both the conversation and the event. Is it possible to discern the structure of a media event from the conversation that is taking place around it?

With my Yahoo! Research colleagues Lyndon Kennedy and Elizabeth F. Churchill, we have begun to examine the relationship between the media event and the impact the event has on the structure and content of the concurrent but remote conversations that it engenders. This is the media's conversational shadow: the backchannel of social conversation cast from a shared event context. The goal is to attain a set of targeted measures through which we will be able to predict audience behaviors as events unfold. This should inform us how to design for participation—which should begin to blur the distinction between the event and its shadow. Before we can dive in, we have to understand the underlying structure and description of the conversation.

back to top  Talking Through the Movie

The primary intuition, which is derived from human-centered social TV research [1], is simple. People stop with the conversation when something important happens—the hush at the start of a movie or that gasp at a missed goal. We have seen the same pattern in online chat rooms with shared video. Viewers tend not to communicate with each other while the video is playing, often saving conversations for logical breaks or lulls. Which is to say, the directed conversation comes to a stop. Presumably, when you do this, you have something to say which takes priority over the event.

A while back, I collected a very small handful of tweets from the first 2008 U.S. presidential debate. I wanted to see if people tweeted more after the debate, so I plotted the number of tweets per minute. There was a three-fold increase right when the debate stopped!

Even more interesting, the volume of tweets per minute during the debate was rather periodic and sinusoidal, which made me wonder if the contour hinted at the start and stop of each debate question segment. I used Newton's Method to find the roots of the graph. Then, using a seven-minute sliding window, I selected only the extrema greater than 1.5 standard deviations from the mean. The resulting set of minutes work like "cuts" as question topic boundaries with about 92 percent accuracy [2]. All this is done without looking at any text.

The question then became, "Can this method of 'naïve tweet counting' scale?"

back to top  Twitter's Explosion

Twitter has grown at an incredibly fast rate. In January of 2009, CNN Breaking News had around 86,000 followers. Nine months later, its follower count had exceeded 2.7 million. Twitter began offering data-mining push feeds that hands back 600 tweets a minute. Track feeds push all the tweets related to a search, and Twitter's "Firehose" can deliver everything if you can get your hands on it. I previously sampled 3,000 tweets from a 90-minute debate. I saw 52,000 tweets from a 90-minute sample of President Obama's inauguration speech.

Addressing scale, I revisited the human-centered observation I originally questioned at Dolores Park. If we position people tweeting while watching a live event, either in person or on television, conversation is represented as talking to someone else: in effect directing a tweet with the @ symbol (using "@" before a username like @barackobama pushes a highlighted tweet into that person's feed). The highlighted tweet becomes more visible; it calls attention to the user, like tapping your friend on the shoulder in a crowd.

Aggregating the number of @ symbols per minute finds these lulls and swells in conversation. In the case of Obama's inauguration, on each minute, 25 percent of the tweets had an @ symbol with the exception of 3 consecutive minutes when the swearing-in took place. What we did not expect was that the number of tweets about the inauguration grew while the number of @ symbols dropped. See Figure 1.

Unlike Newton's Method across aggregated volume this works at Twitter's current scale and is more robust against different sampling techniques. Moreover, the more @ symbols we see on any given minute, the less important that moment should be.

back to top  A Table of Contents

Finally, when we turn to examining the text itself, a table of contents emerges. For this pick your favorite off-the-shelf information retrieval tool; mine is TF/IDF. What we want is a very specific table of contents, so we split the 90-minutes into 5-minute chunks, then find a highly salient slice term that is not salient in the other chunks. Looking back at the inauguration, this produces topic segments like: booing, Aretha, Yo-Yo Ma, anthem, and so on. Couple these words with our importance proxy, and we find the illuminating moments of Aretha Franklin's performance, Obama's speech, and the nation anthem to have greater importance than the other moments.

The signal of human activity here is clean and simple. Our analysis was done on a MacBook Pro using R and MatLab, without the need for large-scale assistance from a Hadoop cloud or Mechanical Turk.

Tweets can be used in social-multimedia research, when facilitated through human-centered research, to identify the shape of the related event. More than just text alone, the form of the conversational shadow has to account for the structure of tweets themselves. Hashtags, @ mentions, and other communication mechanism, both structured and ad-hoc, follow but likely do not mirror the visual content of the event itself. This link between two disparate data streams (online conversations and live events) provides rich opportunities for further investigation.

As we explore new approaches to better navigate, communicate, visualize, and consume events, our tools and research should firmly be based in the conversation structure, be it Twitter, Facebook, or Flickr, and be motivated by our interactions in everyday life, less we miss the game in the park.

back to top  References

1. Cesar, P., Geerts, D., and Chorianopoulos, K. Social Interactive Television: Immersive Shared Experiences and Perspectives. Information Science Reference, 2009.

2. Shamma, D. A., Kennedy, L., and Churchill, E. F. Tweet the debates: Understanding community annotation of uncollected sources. In WSM '09: Proceedings of the international workshop on Workshop on Social Media [Beijing, China, 2009], ACM.

back to top  Author

David Ayman Shamma is a research scientist in the Internet Experiences group at Yahoo! Research. He researches synchronous environments and connected experiences both online and in the world. He designs and prototypes systems for multimedia-mediated communication, and develops targeted methods and metrics for understanding how people communicate online in small environments and at web scale. Ayman is the creator and lead investigator on the Yahoo! Zync project.

back to top  Footnotes

DOI: http://doi.acm.org/10.1145/1869086.1869099

back to top  Figures

F1Figure 1. The fluctuations in volume of Twitter messages sent during U.S. President Barack Obama's inauguration show moments in time when users paid more attention to the ceremony [drop in volume] than to their computer use.

back to top 

©2010 ACM  1528-4972/10/1200  $10.00

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee.

The Digital Library is published by the Association for Computing Machinery. Copyright © 2010 ACM, Inc.


There are no comments at this time.


To comment you must create or log in with your ACM account.