Recognize user intents in online interactions from massive social media data

Online interactions, especially user generated contents on social events, reveal a variety of communicative purposes ranging from expressing feelings to proposing suggestions. Recognizing intents in users' online interactive behavior from massive social media data can effectively identify users' motives and intents behind communication and provide important information to aid monitoring, analysis and decision making for a variety of applications. However, user intents recognition from online communication is inherently challenging due to the ambiguity in semantic processing and diversity of syntax expressions. Moreover, the massive online data are usually unlabeled, which greatly hinders the usage of typical machine learning based methods that can automate the recognition process. In this paper, we tackle this problem by proposing a Speech Act Theory guided classification scheme, which regards online communication as performative actions of users and classifies user utterances according to their pragmatic functions. On the basis of this, we construct a dictionary of performative words, expand it using external knowledge sources and refine it by word embedding and similarity comparison. We then use this dictionary to automatically label the online textual data with intents. With a large amount of the labeled data, we train feature based classifiers to identify user intents in their online interactions. An experimental study using a microblog dataset on social events from SinaWeibo shows the effectiveness of our proposed method.