Automatic classification of orders lines in joint Dealer Management Systems

Abstract

In this thesis we looked at different data originating from several Dealer Management Systems. By comparing the different data we tried to find a field set that can be used as features for our classifier models of receipts. We found that this data can be uniformed well by taking the few fields an intersection of the fields of all Dealer Management Systems yield. When we add some extra fields with slight manipulations we created a data set that has high potential for machine learning classifications. Different set ups showed F1 scores for classification well above 90% through three data sets with four learning models. Further we introduce new options in attempt to improve the classification rate further. We used our domain knowledge for the construction of smart token detectors and construct a unique compound word splitting algorithm for splitting Dutch compound words.