%0 Generic
%9 ['eprint_fieldopt_thesis_type_M.A.' not defined]
%A Daub, Johannes
%C Heidelberg
%D 2021
%F heidok:30794
%R 10.11588/heidok.00030794
%T Tool Support for the Automatic Analysis of Natural Language User Statements
%U https://archiv.ub.uni-heidelberg.de/volltextserver/30794/
%X [Context & Motivation]  Developers need to learn about the requirements of software users, who give their feedback mostly in form of natural language statements. Processing these statements through manual coding, however, is an elaborate task and makes it unsuitable for big datasets. By extracting concepts from these statements, developers can get insights about the point of view of the software user. A software tool that provides automatic processing can help with this process.    [Contributions]  This thesis explores the state-of-the-art topic modeling methods for user forums and applies suitable methods in the context of concept detection to a manually collected and annotated interview dataset. A software tool for automatic language processing, named "Feed.UVL" is created and the selected methods are integrated into this tool. The created software tool provides dataset management, which means that datasets can be stored, reviewed and deleted with the software. The implemented methods can be used to analyze these datasets for concepts. With the result visualization, the analysis results can be reviewed and the performance can be evaluated via the F1-score on a ground truth. Feed.UVL uses a micro-service architecture, which means it can be extended easily with new methods or functions. The integrated methods are then evaluated for the task of concept detection. A set of quality assurance measures, including static code analysis, component and system tests, have also been performed on the created tool.    [Conclusion]  The main part of the thesis was the creation of a novel tool for natural language processing. The tool has a clean and user-friendly design and supports researchers in their analysis. Automatic analysis tasks can be handled and the user interface provides a rich display of results, including the metrics false positives, false negatives, precision, recall and F1-score. The current design and micro-service architecture ensures that the tool can be extended easily for further analysis methods and future research goals. At the moment, two state-of-the-art topic modeling methods (LDA and SeaNMF) are integrated, which were adapted for the use in concept detection. The evaluation has shown that while their precision is relatively high (0.84 for LDA and 0.83 for SeaNMF), their recall is rather low compared to a manually annotated ground truth for use in concept detection, which leaves space for improvements and future works.