LOCAL EXPLAINABILITY FRAMEWORK FOR NLP BASED CLASSIFICATION SYSTEM

Know More

CLIENT

A global provider of financial market data and infrastructure

BUSINESS CHALLENGES

The need was to conduct automated enhanced due diligence of entities (people, organizations) from multiple sources and create a complete profile with information like ownerships, beneficiaries, negative news, and so on. Classification of negative news was an important aspect of this report. Wrong classification of information impacted the quality of reports generated for client’s customers.

SOLUTION

A natural language processing-based classification system was developed to collect news about a specific entity and check if there were any negative news about the person or the organization. The classification system -

Categorizes the news into 8 different classes based on the content (e.g., bankruptcy, fraud, criminal etc.)
Highlights the keywords in the content using a local explainability framework. These keywords are used by the models to do the classification.

BUSINESS BENEFITS

Achieved accuracy greater than 85% in classifying the items in different negative news classes along with the confidence score of the classes.

Ability to determine if the predicted class for the low confidence instances were correct and update the wrong ones.

Improved the productivity of the associates by 30%, enabling them to focus on high value research activities.