What kind of #conversation is Twitter? Mining #psycholinguistic cues for emergency coordination

• Information filtering model to reduce Twitter traffic for disaster coordination.
• Modeling of coordination via conversation using pyscholinguistic theories.
• Domain-independent linguistic cues distinguish likely conversation up to ROC = 0.84.
• Higher information density in positively classified conversational messages.
• Proof of same human behavior in face-to-face and online (mediated) communication.

The information overload created by social media messages in emergency situations challenges response organizations to find targeted content and users. We aim to select useful messages by detecting the presence of conversation as an indicator of coordinated citizen action. Using simple linguistic indicators drawn from conversation analysis in social science, we model the presence of coordination in the communication landscape of Twitter1 using a corpus of 1.5 million tweets for various disaster and non-disaster events spanning different periods, lengths of time, and varied social significance. Within replies, retweets and tweets that mention other Twitter users, we found that domain-independent, linguistic cues distinguish likely conversation from non-conversation in this online form of mediated communication. We demonstrate that these likely conversation subsets potentially contain more information than non-conversation subsets, whether or not the tweets are replies, retweets, or mention other Twitter users, as long as they reflect conversational properties. From a practical perspective, we have developed a model for trimming the candidate tweet corpus to identify a much smaller subset of data for submission to deeper, domain-dependent semantic analyses for the identification of actionable information nuggets for coordinated emergency response.