Semantic AI Enables Multi-channel Tax Advisory Innovation

The organization employed an army of individuals who manually researched and provided knowledge about changes and updates in tax law, which was costly in time, labor and increased corporate risk. When searching for relevant information, tax advisors were required to look in multiple places and struggled to find the most current version. To improve operations and provide their clients with up-to-date information, they needed an automated way to harmonize information, determine the context of a piece of content and leverage the context to drive a fast and accurate search and navigation experience.

Project goals:

• Establish a centralized database that contains all available tax knowledge – everything a tax advisor needs to provide quality services to their clients
• Improve advisor efficiency and the quality of research results on the internal search portal
• Digitize the content creation process using smart content and tooling
• Create a real-time personalized view of news and legislation via the news app
• Eliminate manual processing and automate the creation of metadata and relationships between content

They chose Semaphore’s Semantic AI platform Semaphore to enable fast and accurate search so users can find what they need when they need it.

To begin the process they leveraged Semaphore’s Knowledge Model Management (KMM) capabilities to build a high level, flat model that contains concept schemes, topics, subscriptions and weekly stream information associated with tax advisory. The model contained a robust set of alternative labels, which are essential to enhance the classification process.

Semaphore Classification and Language Services (CLS) automatically examines each information asset and appropriately tags and organizes it by content type (i.e. journal, court, government) and context. To support the news app, each concept contains a production and score indicator. When classification results produced a score of 70% or greater, the article is pushed to the news app. News items are added to a user’s weekly stream, which is organized by week and available immediately.

They created a document fingerprint using the FACT extraction framework that allows them to automatically examine their more than 50M files and identify the document type (i.e. case law, internal advice …) and context. This allows them to automatically determine what a piece of content is about in a precise, consistent and reliable manner.

By creating a feedback loop to enrich the model, users can provide input when they find misclassified articles/content via the mobile app. Feedback is reviewed by model managers and if misclassification is confirmed, the model is enriched with the new concept/information. While they expected to receive only negative classification responses, they were pleasantly surprised to receive positive (70%) as well as negative (30%) classification responses.