# Desired corpus content
Product-centric. For a given product:
 - Reviews
 - Images. Limit resolution and angles somehow? Don't need
   thumbnails, for sure.
 - Description(s)
 - Ratings
 - Info for all styles
# Desired corpus layout
For researchers, something simple, like Tamara Berg's attribute
discovery dataset --- a hierarchical directory structure
(keeping a manageable number of files in each leaf directory)
with task-specific items for each product, organized by file
name.