Crowd-Powered Replies to Public Twitter Questions

“Friendsourcing” information seeking by posting questions to a social networking site, like Twitter, is becoming increasingly common. However, unlike conventional search engines, friendsourced information seeking does not always return an answer. Examining over 85,000 information-seeking tweets from the public Twitter firehose, we found that only about a third received replies at all.

We created a Twitter agent that monitors the public feed for information-seeking questions, generates an answer via crowdsourced labor, and tweets the reply back to the original asker. This image shows an example:

Our crowd-powered Twitter agent serendipitiously offers advice about how to clean Sperrys (a type of shoe). The asker, pleased with the response, favorited the reply and retweeted it.

Our crowd-powered Twitter agent serendipitiously offers advice about how to clean Sperrys (a type of shoe). The asker, pleased with the response, favorited the reply and retweeted it.

Raters evaluating naturally occurring “friendsourced” answers to questions on Twitter and the answers generated via our crowdsourcing project rated them as being of equally high quality. Indeed, the people who received replies from our agent were generally quite pleased with the experience — over a third of them thanked our agent in some way, such as by favoriting the tweet or sending a reply message on Twitter.

Our work illustrates ways in which a “socially embedded search engine” can augment basic social network Q&A experiences.

For more, see our full paper, A Crowd-Powered Socially Embedded Search Engine. You can also view some sample Q&A exchanges our agent participated in at our project’s Twitter page.

Jin-Woo Jeong (Hanyang University), Meredith Ringel Morris, Jaime Teevan, & Dan Liebling (Microsoft Research)

4 thoughts on “Crowd-Powered Replies to Public Twitter Questions

  1. This is very cool work!

    How much overlap was there in the question set? Could you build up a set of answers over time and reuse them?

  2. Possibly, for certain types of factual questions. However, many questions were quite personalized (e.g., should I dye my hair blonde or brown?), and so answers for those ones wouldn’t really transfer across users.

  3. For these personal questions like dying your hair, then, was the “crowdsourced labor” able to provide useful responses? Seems like that would be tough.

Comments are closed.