1080 Examples
Here are some experiments on an even larger collection of documents. There are included less because they are successful examples of Explainers, but more because they show how the Explainers diagrams work (or fail) at scale. I cannot give out the data for these (although I am working on creating a new version of this data that is shareable).
Warning: these explainers are HUGE SVG files (multiple megabytes) - and they may make your browser choke.
For these experiments, the data set is a collection of 1080 documents from the period 1520-1800 taken from the "EEBO/TCP" document collection. We randomly sample from the document collection such that we get the same number of documents for each decade (details elsewhere).
In the first experiment, we try to identify "verse" (or poetry). There were genre tags for "verse" and "prose" (which is clearly not verse). There are also genre tags like "poetry" and "legal documents." For training, we only considered the documents labeled "verse" and "prose" and trained an explainer to distinguish these. In the diagrams, we show all documents - coloring the ones in the "verse" and "prose" classes, as well as the ones in the "poetry" classes. As we would hope, poetry looks pretty similar to verse.
(link to experiment - it's too big to embed)
In this second experiment, we try to create an Explainer for date. We train on "was it written before 1680" vs "was it written after 1710" (note that we have a 30 year buffer where we don't classify things either way).
You can notice that, in general, older documents are more likely to be scored as being written before 1680 (phew). There are a few outliers that were way ahead of their time (the white boxes near the bottom). To give a sense of what the "right answer" would look like, the rightmost diagram shows the "ground truth" - what would happen if we used the date written as the score.
