FamilyAudit-Classifier

A Multiclass sequence classifier based on BERT base italian, fine-tuned on selected corpora from Municipalities's Family Audit plans

AIxPA

• kind: product-template
• ai: NLP
• domain: PA

The context in which this project was developed: The classifier is trained to suggest one or more labels within the Family Audit framework. More specifically, the classifier can predict the category of the action ("azione") of the text describing it. Given that the category of action for Municipalities is univocally connected with one macrocategory ("macrocategoria") and one field ("campo"), this classifier can be used to indicate which macrocategory and field of the action the text belongs to.

The classifier is a tool that can be actually used for any classification domain, given similar or bigger amounts of text data. It does not matter which are the labels to predict, but they need to be encoded as integers (e.g. "Ecology" is represented by 1, "Childcare" is represented by 2, etc.). The product contains operations for

• Log the data in order to prepare it for training
• perform model training and registering the model
• serving the model using a custom API.
• collect the data using monitoring gateway.

Usage

Tool usage documentation here.

How To

• Preprocess data for training
• Train the classifier model
• Expose the classifed model as a service
• Collect production data and metrics

License

Apache License 2.0