Bridging Research Fields: An Empirical Study on Joint, Neural Relation Extraction Techniques

Abstract

Information systems that have to deal with natural language text are often equipped with application-specific techniques for solving various Natural Language Processing (NLP) tasks. One of those tasks, extracting entities and their relations from human-readable text, is relevant for downstream tasks like automated model extraction (e.g. UML diagrams, business process models) and question answering (e.g. in chatbots). In NLP the rapidly evolving research field of Relation Extraction denotes a family of techniques for solving this task application-independently. Thus, the question arises why scientific publications about information systems often neglect those existing solutions. One supposed reason is that for reliably selecting an appropriate technique, a comprehensive study of the available alternatives is required. However, existing studies (i) cannot be considered complete due to irreproducible literature search methods and (ii) lack validity, since they compare relevant approaches based on different datasets and different experimental setups. This paper presents an empirical comparative study on domain-independent, open-source deep learning techniques for extracting entities and their relations jointly from texts. Limitations of former studies are overcome (i) by a rigorous and well-documented literature search and (ii) by evaluating relevant techniques on equal datasets in a unified experimental setup. The results$^{1}$ show that a group of approaches form a reliable baseline for developing new techniques or for utilizing them directly in the above mentioned application scenarios($^{1}$Our code and data: https://github.com/JulianNeuberger/re-study-caise.).

Our work is supported by the Bavarian Research Foundation (grant no. AZ-1390-19).