Data Driven Creation of Sentiment Dictionaries for Corporate Credit Risk Analysis

It has been shown, that German-language user generated content can improve corporate credit risk assessment, when sentiment analysis is applied. However, the approaches have only been conducted by human coders. In order to automate the analysis, we construct 20 domain-dependent sentiment dictionaries based on parts of a manually classified corpus from Twitter. Then, we apply the dictionaries to the remaining part of the corpus and rank the dictionaries based on their accuracy. Results from McNemar’s tests indicate, that the three best dictionaries do not differ significantly, but significant difference can be assured regarding the first and the fourth dictionary in the ranking. In addition to that, a general German-language dictionary is inferior compared to the constructed dictionaries. The results emphasize the importance of domain-dependent dictionaries in German-language sentiment analysis for future research. Furthermore, practitioners can utilize the dictionaries in order to create an additional indicator for corporate credit risk assessment