Extraction of semantic relations in noisy user-generated law enforcement data

Relation extraction from text is a well-known and extensively studied topic in Natural Language Processing research. However, the implementation of relation extraction approaches in real-world application scenarios raises various methodological considerations which are often left implicit in existing research. This paper explores these considerations using a real-world dataset of user-generated police reports in Dutch. The use of linguistic features based on dependency trees is investigated, including an ablation analysis of the importance of individual features. The construction of negative examples for machine learning models is discussed, as well as the construction of a baseline model. The methodological implications of using a small dataset are discussed in terms of the design and performance of a Long Short Term Memory network as well as a Support Vector Machine. In general the models perform well, however the definition of the classification task, and in particular the construction of negative examples, are shown to have a large impact on classification accuracy and subsequently on the interpretation of the evaluation results.