Creating Sentiment Dictionaries: Process Model and Quantitative Study for Credit Risk

Abstract

Since textual user generated content from social media platforms contains valuable information for decision support and especially corporate credit risk analysis, automated approaches for text classification such as the application of sentiment dictionaries and machine learning algorithms have received great attention in recent user generated content based research endeavors. While machine learning algorithms require individual training data sets for varying sources, sentiment dictionaries can be applied to texts immediately, whereby domain specific dictionaries attain better results than domain independent word lists. We evaluate by means of a literature review how sentiment dictionaries can be constructed for specific domains and languages. Then, we construct nine versions of German sentiment dictionaries relying on a process model which we developed based on the literature review. We apply the dictionaries to a manually classified German language data set from Twitter in which hints for financial (in)stability of companies have been proven. Based on their classification accuracy, we rank the dictionaries and verify their ranking by utilizing Mc Nemar's test for significance. Our results indicate, that the significantly best dictionary is based on the German language dictionary SentiWortschatz and an extension approach by use of the lexical-semantic database GermaNet. It achieves a classification accuracy of 59,19 % in the underlying three-case-scenario, in which the Tweets are labelled as negative, neutral or positive. A random classification would attain an accuracy of 33,3 % in the same scenario and hence, automated coding by use of the sentiment dictionaries can lead to a reduction of manual efforts. Our process model can be adopted by other researchers when constructing sentiment dictionaries for various domains and languages. Furthermore, our established dictionaries can be used by practitioners especially in the domain of corporate credit risk analysis for automated text classification which has been conducted manually to a great extent up to today.