Detecting Lexically Cohesive & Temporally Bounded Tweet Sessions on Twitter Timelines
Hayes, Brittany · Carolina Digital Repository (University of North Carolina at Chapel Hill) · 2019
This exploratory study examines the concept of the "tweet session" - instances where a user of the microblogging site Twitter posts two or more related tweets in a short period of time. The study outlines a set of characteristics that aid in the detection of topically cohesive units. Four of these properties are external to the tweet text: a 24-hour time frame, inclusion of at least two tweets, inclusion of originally authored tweets and the exclusion of replies. Four additional properties were derived from the natural language processing and information retrieval literature: lexical cohesion based on unigram and character bigram feature representations, conjunction use, signals of continuation, and anaphora resolution. A sample of 220 user timelines was analyzed to detect series of tweets meeting the definition of a session as conceptualized in the study. 93.6% of timelines included at least one technical tweet session. Lexical cohesion as determined by cosine similarity retrieved the most sessions: 815 technical sessions when unigrams were used as the unit of tokenization, and 1391 technical sessions when character bigrams were used. The majority of users engaged in tweet sessions exceeding 140 characters (the Twitter-imposed limit for a single tweet); however, when unigrams were used in the feature representation approximately 47% of timelines had tweet sessions of less than 140 characters on average. This research shows that tweet sessions exist and can be detected by computational means.