Big 4 Consulting Company

Regulatory changes and judgements impact the tax guidance issued for clients. Traditionally, this guidance is compiled by experts at one point in time. A first step towards "always-current" advice is to automatically surface news items that impact client-specific issues, from a wide variety of high-volume data sources. Using algorithmic natural language understanding, can we automate the identification of relevant information?

Our system uses REST API queries to autonomously query news feeds. When new items are posted, they are downloaded and compared against all existing guidance and tax memos. We decided to use a hybrid scoring approach for our surfacing engine. We first built a custom knowledge graph derived from federal tax code, current treasury regulations, and other metadata. This involved automated PDF text extraction of tens of thousands of documents, as well as various web crawling tasks. We then combined the context provided by the knowledge graph with features derived via natural language processing (topic modeling) from the full text of news items. We validated this hybrid similarity scoring model using ground truth provided by tax experts. Only new items that are relevant will alert the tax practitioner and be presented for review.

The prototype we designed successfully demonstrated feasibility of our approach, and we created a follow-on minimum viable product. The system fundamentally acts as a highly accurate filter that understands both language and context. Our approach enables this consultancy to start moving from "point-in-time" to continuously updated client guidance, creating new revenue streams.