Feedback are user approved or corrected extraction values that are submitted by you back to the extraction system.
You should always submit the feedback on extractions. The feedback that is correctly submitted allows Gini to measure the extraction quality in order to provide the best possible user experience. Moreover, it helps us to continuously improve the extraction system and perform valuable data analytics for you.
The feedback should only be sent if real users can see, approve, correct and complement the extractions based data.
Gini employs various machine learning techniques in order to learn from feedback automatically. Therefore, it is equally important for Gini to receive both feedback on correct and on incorrect extractions.
There are currently two ways to submit the feedback.
The first and the most common one is to submit the complete feedback in one request. This is the easiest way if your frontend (application) displays the extractions on a screen in an editable form. A user can modify the extractions before pressing the confirmation button.
The second (a rare use case) is where the final approvement signal (button click) is not possible. In such case you can send the feedback on one label per request.
There are three different types of feedback:
positive feedback: (our preferred type) the extraction was correct and confirmed by the user without modification.
complementary feedback: The extraction system extracted nothing, e.g. the response does not contain the requested label. The user manually entered the correct value.
negative feedback: The extraction was incomplete or erroneous. The user manually corrected the extraction.
informative feedback: The extraction was changed for another reason. For example if a line item is returned, the quantity extraction gets feedback "0" and a return reason.
Please see detailed examples next.
The Gini Pay API allows you to submit the feedback on multiple extractions for a single document with a single request. It is strongly recommended that you submit your feedback in this way for two reasons. On the one hand, the total number of round trips is reduced to one and the feedback is handled internally as a batch. Therefore, the update is more efficient for multiple extractions compared to submitting the feedback with each separate request (see single feedback). On the other hand, Gini's machine learning training techniques can benefit from the feedback on multiple extractions since Gini will be aware of the fact that single parts of the submitted feedback belong together.
You should send the feedback only on those labels that has been physically seen by the user. Unseen labels should be filtered out.