Applying Deep Learning to Examine Tax Footnotes: A Study of Emotions and Tax Outcomes

This study applies deep learning algorithms provided and validated by the International Business Machine (IBM) Watson to examine the emotional content of tax footnotes prepared by the corporation. Employing IBM Natural Language Understanding (NLU) application programming interface (API) and IBM Tone Analyzer Service, I provide evidence that emotions detected in tax footnotes are associated with firms’ tax risk, which is captured by the magnitude of unrecognized tax benefits (UTBs). Also, I show that firms change their sentiment, not emotion, in tax footnotes disclosure after the Internal Revenue Service (IRS) required the filing of Schedule UTP in 2010, implying that, unlike emotion, sentiment (i.e., a positive, negative, or neutral tone) is easier to modify. Furthermore, I demonstrate that emotions detected from tax footnotes can predict tax cash paid and UTBs related settlements with taxing authorities, respectively. Utilizing artificial intelligence (AI) technology and sophisticated natural language processing (NLP) technique, I take textual analysis one level deeper to better understand the growing volume of unstructured content.

Many thanks to Tony’s Dissertation Committee: o Committee Chair: Dr. Anthony Curatola - Professor - Drexel University o Committee Member: Dr. Helen Choy - Associate Professor - Drexel University o Committee Member: Dr. Hsihui Chang - Professor - Drexel University o Committee Member: Dr. Xin Dai - Assistant Professor - Drexel University o Committee Member: Dr. Vadake Narayanan - Professor - Drexel University