Measuring, Predicting and Visualizing Short-Term Change in Word Representation and Usage in VKontakte Social Network

Language in social media is extremely dynamic: new words emerge, trend and disappear, while the meaning of existing words can fluctuate over time. Such dynamics are especially notable during a period of crisis. This work addresses several important tasks of measuring, visualizing and predicting short term text representation shift, i.e. the change in a word's contextual semantics, and contrasting such shift with surface level word dynamics, or concept drift, observed in social media streams. Unlike previous approaches on learning word representations from text, we study the relationship between short-term concept drift and representation shift on a large social media corpus - VKontakte posts in Russian collected during the Russia-Ukraine crisis in 2014-2015. Our novel contributions include quantitative and qualitative approaches to (1) measure short-term representation shift and contrast it with surface level concept drift; (2) build predictive models to forecast short-term shifts in meaning from previous meaning as well as from concept drift; and (3) visualize short-term representation shift for example keywords to demonstrate the practical use of our approach to discover and track meaning of newly emerging terms in social media. We show that short-term representation shift can be accurately predicted up to several weeks in advance. Our unique approach to modeling and visualizing word representation shifts in social media can be used to explore and characterize specific aspects of the streaming corpus during crisis events and potentially improve other downstream classification tasks including real-time event detection.