Tax Disclosure Text Clustering
Description
This project takes written data from uncertain tax disclosures, cleans up errors in the text, incorporates additional attachment information not present in the Compliance Data Warehouse (CDW), and groups the disclosures into semantically similar topics. This creates a single source to find uncertain tax disclosure declarations and additional information to facilitate review. The process follows three steps: 1) extraction of text from Form 8275-R, Regulation Disclosure Statement, and Schedule UTP, Uncertain Tax Position Statement, from CDW and supplemental attachments, 2) embedding of text into numeric vectors using the E5-large model, and 3) performing hierarchical clustering on the text vectors to group the disclosures.
Detailed example
Written descriptions provided by taxpayers disclosing items or positions not otherwise adequately disclosed on a tax return to avoid certain penalties are grouped into semantically similar topics to create clusters. The model produces a table of cleaned tax disclosures and their assigned cluster (numerical label).
AI / analytics pattern
Natural Language Processing: AI that processes, interprets, and shares information in human language.
Automation level / stage
c) Deployed – The use case is being actively authorized or utilized to support the functions or mission of an agency.
Expected benefit
The primary goal of this use case is to gain insights useful for improving future examinations and outcomes.
Audit / financial statement impact
The output is not presumed to be high-impact and is not used as the principal basis for significant decisions/actions
Controls / human review
ATO: Yes; PIA: Not published
Data needed
Hierarchical Clustering determined by cutting the dendrogram at a user specified height, which in this case represents the minimum cosine distance at which clusters are determined to be distinct. This consisted of evaluation by the subject matter experts (SMEs). One critical aspect to the evaluation involved setting a minimum similarity threshold for the clustering process. We also conducted spot checks on a random sample of groups to ensure consistency.