Build Your Model

📙 Quick introduction to some key terms before we get to the actual training process in Jasper.

Object Detection

Traditionally Optical Character Recognition (OCR) is used for document processing to extract text content from scanned images. The approach works well for detecting and transcribing the text in the images. The transcribed text can be pushed to any natural language processing models to perform classification, named entity recognition and more. This approach fails when there is complex document with lots of formatting, lines, tables, and fields, take the case of a bank statement or 1040. OCR won’t preserve the formatting thus makes it difficult to find the regions or areas of interest. In such cases breaking down the image to sub images of useful area will come handy. This process can be achieved using the computer vision technique, object detection.

Jasper combines the power for of deep learning models for Object detection and custom OCR for document extraction to handle complex tables, forms , check boxes etc.

Named Entity recognition (NER)

Named Entity Recognition is a process where an algorithm takes a string of text (sentence or paragraph) as input and identifies words mentioned in the text to relevant categories like (people, places, and organizations) that are mentioned in that text. Some of the use cases include Classifying content for news providers, Chat bots etc. In the context of Jasper we use NER models to extract content from Legal Contracts, Resumes and any text content.

Training in Jasper

Step 1 - Annotate your training data

Your model is as good as your data. Make sure you have annotated a good number of samples which covers all the variations that could potentially come across in the production scenarios

Step 2 - Save all your labels

Make sure you hit the 'Save' button in the annotation interface

Step 3 - Start the training process

Go back the document listing page and trigger the 'Train' button. If you are training this project for the first time , the status will be in 'Labeling'

You will asked to pick the document type to be trained. Pick the right one from the options.

Training a model is time a time consuming process. The time to finish depend on your training data volume , labels and the infrastructure (GPU load etc). Once the training is triggered, the project status will change to 'In Training'.

Once the training process is complete the project status will change to 'Trained'

Accuracy Scores

Once the project is trained status , its time to see some key metrics of your model. This is available under the 'View Training Summary'

Accuracy of a ML Models is often expressed in terms of 3 key metrics; Precision, Recall and F1 score.

Precision

Precision is the ratio of system generated results that correctly predicted positive observations (True Positives) to the system’s total predicted positive observations, both correct (True Positives) and incorrect (False Positives).

In other words, precision answers the following question:

How many of those Payee Names labelled by the system as are actually Payee Names?

Recall

Recall is the ratio of system generated results that correctly predicted positive observations (True Positives) to all observations in the actual class (Actual Positives).

In other words, recall answers the following question:

Of all the Payee Names present in the form or scanned document, how many of those did the system correctly classify as Payee Names?

F1 Score

F1 Score is the weighted average (or harmonic mean) of Precision and Recall. This score takes both False Positives and False Negatives into account to strike a balance between precision and Recall. F1 Score might be a better measure vs. accuracy if we need to seek a balance between Precision and Recall AND there is an uneven class distribution