It's hard to find time to post when we're busy. Here's a glimpse of a recent piece of work we're quite excited about. Our form comprehension algorithms have improved in dynamically understanding a form structure as well as extracting user inputs as structured values.

Consider the following manual form (pay no attention to the content, it is purely illustrative). We previously had difficulty in accurately identifying check-boxes, as well as whether these were checked or not.

Updated output: [Note1: First char indicates check "X" vs unchecked "O"] [Note2: Second char is cluster id for grouped form elements] X 1 What is your favorite editor? --> Vim O 1 What is your favorite editor? --> Emacs O 2 Is pineapple OK on pizza ? --> Yes X 2 Is pineapple OK on pizza ? --> No X 3 What are your favorite days of the week? --> Monday O 3 What are your favorite days of the week? --> Tuesday O 3 What are your favorite days of the week? --> Wednesday O 3 What are your favorite days of the week? --> Thursday O 3 What are your favorite days of the week? --> Friday O 3 What are your favorite days of the week? --> Saturday O 3 What are your favorite days of the week? --> Sunday X 4 Oxford comma or normal comma ? --> Oxford O 4 Oxford comma or normal comma ? --> Normal X 5 Star wars is better than Star Trek? --> Yes O 5 Star wars is better than Star Trek? --> No

We're now able to generate fairly accurate output without any special fine-tuning. There are still some hiccups, e.g. the O from Oxford interpreted as a checkbox, but the performance is already quite strong, especially in standard computer-generated forms where handwriting is less of a challenge.