Predicting Temporal Intention in Resource Sharing

When users post links to web pages in Twitter there is a time delta between when the post was shared (t(tweet)) and when it was read (t(click)). Ideally, when this time delta is small there is often no change in the page's state. However upon reading shared content in the past and due to the dynamic nature of the web, the page's state could change and the intention of the author need to be inferred. In this work, we enhance a prior temporal intention model and tackle its shortcomings by incorporating extended linguistic feature analysis, replacing the prior textual similarity measure with semantic similarity one based on latent topic detection trained on Wikipedia English corpus, and finally by enriching and balancing the training dataset. We uncovered three different intention behaviors in respect to time: Stable Intention, Changing Intention from current to past, and Undefined intention. Using these classes and only the information available at posting time from the tweet and the current state of the resource, we correctly predict the temporal intention classification and strength with 77% accuracy.