Comparison of Lexicon Performances on Unstructured Behavioral Data

Our challenge is gaining insight, analyzing and understanding implicit behavior behind text. Useful tools are machine learning and sentiment analysis. Text classification copes with content mass to extract actionable knowledge. We evaluate relative performance of four well-known lexicons implemented through Nave Bayes and bigram analysis. Results show Loughran Lexicon captures well predatory practices. Nave Bayes modeling provides means to compare lexicons. We obtain 87 % accuracy with Loughran. Cross Validation validates prediction accuracy. We support our hypothesis that lexicons are not sufficient for AI tools. Lexicons allow easier feature selection for Nave Bayes approach. Our goal is to understand sentiment and emotion behind complaints and provide clearly relevant training data as one attempts to automate decision making using AI tools.