Sub-event detection from twitter streams as a sequence labeling problem

Proceedings of NAACL-HLT 2019, pages 745–750 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 745 Sub-Event Detection from Twitter Streams as a Sequence Labeling Problem Giannis Bekoulis Johannes Deleu Thomas Demeester Chris Develder Ghent University - imec, IDLab Department of Information Technology firstname.lastname@ugent.be Abstract This paper introduces improved methods for sub-event detection in social media streams, by applying neural sequence models not only on the level of individual posts, but also di- rectly on the stream level. Current approaches to identify sub-events within a given event, such as a goal during a soccer match, essen- tially do not exploit the sequential nature of so- cial media streams. We address this shortcom- ing by framing the sub-event detection prob- lem in social media streams as a sequence la- beling task and adopt a neural sequence archi- tecture that explicitly accounts for the chrono- logical order of posts. Specifically, we (i) es- tablish a neural baseline that outperforms a graph-based state-of-the-art method for binary sub-event detection (2.7% micro-F1 improve- ment), as well as (ii) demonstrate superior- ity of a recurrent neural network model on the posts sequence level for labeled sub-events (2.4% bin-level F1 improvement over non-se- quential models). 1 Introduction Social media allow users to communicate via real- time postings and interactions, with Twitter as a notable example. Twitter user posts, i.e., tweets, are often related to events. These can be so- cial events (concerts, research conferences, sports events, etc.), emergency situations (e.g., terrorist attacks) (Castillo, 2016), etc. For a single event, multiple tweets are posted, by people with vari- ous personalities and social behavior. Hence, even more so than (typically more neutral) traditional media, this implies many different perspectives, offering an interesting aggregated description. Given this continuous and large stream of (likely duplicated) information in Twitter streams, and their noisy nature, it is challenging to keep track of the main parts of an event, such as a soc- cer match. Automating such extraction of differ- ent sub-events within an evolving event is known as sub-event detection (Nichols et al., 2012). For tracking each of the sub-events, the timing aspect is an important concept (i.e., consecutive tweets in time). Thus, a sequential model could successfully exploit chronological relations between the tweets in a Twitter stream as an informative feature for sub-event detection. Several methods have been proposed for sub- event detection: clustering methods (Pohl et al., 2012), graph-based approaches (Meladianos et al., 2015), topic models (Xing et al., 2016) and neural network architectures (Wang and Zhang, 2017). None of these studies exploits the chronological relation between consecutive tweets. In contrast, our work does take into account that chronolog- ical order and we predict the presence and the type of a sub-event exploiting information from previous tweets. Specifically, we (i) propose a new neural baseline model that outperforms the state-of-the-art performance on the binary classi- fication problem of detecting the presence/absence of sub-events in a sports stream, (ii) establish a new reasonable baseline for predicting also the …

Content truncated to 500 words