Also, it's very important that there be the option to see completely random samples, not prioritized samples like what we see here. A generic type of study will be to ask 'how often is word X used in context Y' where context Y can only be determined by a human. For all these settings, one will want to take a few random samples for each time point and hand-annotate them so as to be able to determine, given X at time t, the likelihood that the context is Y; which you could multiply by the raw frequency of X over time to estimate the freq of X in context Y over time. This is a hugely important thing to enable.
Says Erez