Employee Retention Credit Text Clustering
Description
Application of Line-Item Consolidation algorithm to identify highly similar text and apply additional features to similar clusters to enable identification of suspicious texts.
Detailed example
Detailed explanations from taxpayers as to why they are claiming the Employee Retention Credit (ERC) are grouped into semantically similar topics to create clusters. The ERC is a refundable tax credit for certain eligible businesses and tax-exempt organizations that had employees and were affected during the COVID-19 pandemic. 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
This enables the agents to increase their capacity and reduce time researching and identifying potential suspicious trends of topics used in claims.
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
We the effectiveness in the algorithm in supporting the identification of potentially risky clusters. We did this in two ways, the first was to identify clusters that had a high proportion of additional risk flags, since similar text indicates that text is potentially coming from a single source group, the rest of the cluster is also a target for review. The other method was to identify groups with a large number of Preparer EINs. This indicates that the preparer ID is missing, but they may have a shared source. One critical aspect in the evaluation process involved setting a minimum similarity threshold for the clustering process. We also conducted spot checks on a random sample of groups to ensure consistency, where we found that the text within each group was predominantly consistent, highlighting the algorithms’ ability to group semantically similar descriptions.