Make sure you saw this link for preprocessing first

Custom classification is a two step process:

1. Identify labels and create and train a custom classifier to recognize those labels.
2. Once amazon Comprehend trains the classifier, send unlabeled documents to be classified using that classifier.

Training a Custom Classifier

Using the AWS SDK for Python:

Instantiate Boto3 SDK:

import boto3
client = boto3.client('comprehend', region_name='region')

To create a Classifier:

create_response = client.create_document_classifier( InputDataConfig={ 'S3Uri': 's3://S3Bucket/docclass/file name' }, DataAccessRoleArn='arn:aws:iam::account number:role/resource name', DocumentClassifierName='SampleCodeClassifier1', LanguageCode='en')

To run a custom classifier job:

start_response = client.start_document_classification_job( InputDataConfig={ 'S3Uri': 's3://S3Bucket/docclass/file name', 'InputFormat': 'ONE_DOC_PER_LINE' }, OutputDataConfig={ 'S3Uri': 's3://S3Bucket/output' }, DataAccessRoleArn='arn:aws:iam::account number:role/resource name',
DocumentClassifierArn= 'arn:aws:comprehend:region:account number:document-classifier/SampleCodeClassifier1')

Some other notes

• To train a custom classifier, identify the classes you want to use for classification. For each class, for more accurate training, we recommend at least 50 documents or more for each class.

• When training your classifier, the data must be in a single .csv file. The format of the data will depend onwhich classifier mode you choose.

• After you train the custom classifier, you can analyze documents in either asynchronous (batch) or synchronous operations (real time)

• Multi-class mode supports up to 1 million examples containing up to 1000 unique classes.

• Multi-label mode supports up to 1 million examples containing up to 100 unique classes.