Guided Research Sebastian Moser

Summarization of German Court Rulings

Abstract
German Court Rulings contain three basic parts. First, the tenor summarizes the explicit decision a judge made as weil as who has to pay the costs of the trial. Second, the main body of a judgment is divided into the facts and the reasoning, explaining why the judge came to their conclusion given the specific preconditions. Third, the verdict is enriched by meta-information found in the rubrum. This part contains the facts about the judgment such as previous instances, the normchain and the guiding principle and thus is generally at the beginning of a judgment. The guiding principle shortly summarizes why the specific decision was made and thus is extremely useful for legal practitioners. This high level summarization allows them to quickly assess whether the court ruling is important to their own case.
Not all of those pieces of information are found in all verdicts. Especially the guiding principle is only added by the judge or a legal publisher, if the judgment contains an important legal decision.· A lot of expert knowledge and time is necessary to create this summarization. In this guided research, 1 want to explore if it is possible to automatically summarize a judgment and generate this information.
Automatie summarization systems can be divided into extractive and abstractive approaches. In extrac-tive summarization the text is split up into segments, most commonly sentences, which are then ranked based on a scoring mechanism. Common scoring methods are Topic modeling, Latent Semantic Analysis or today more commonly Neural Networks [1, 2). Abstractive summarization is directly capable of creating new text and not limited to copying parts from the document. The state-of-the-art methods all combine two Neural Networks. An encoder network compresses and selects the most important pieces of infor-mation from the document. The decoder utilizes this compressed representation as well as the previously generated words to generate a new word [2, 3, 4). Extractive and abstractive approaches are sometimes combined, i.e. the Neural Network can generate new words but is also allowed to copy words or segments from the input document [5].
As the generic solution to summarization is still an open research question, there are many different approaches with their own advantages and disadvantages. In this guided research, 1 want to test if the state-of-the-art approaches are applicable to German verdicts and how they need tobe changed based on legal domain knowledge to create better summaries. One specific region of interest is the summarization of long documents, as the encoding of their information is more complicated to do [6). As most summarization methods are tested on shorter, English newspaper articles, it will be interesting to see their performance on a different domain. In this regard, extractive and abstractive approaches shall also be compared. To tackle the length of German verdicts, combined methods might be most applicable with a pre-selection of the most important sentences by an extractive approach. As German verdicts contain many references to other documents, it might be necessary to utilize domain knowledge or use different sources of text to create high-quality summarizations of a verdict.
To evaluate the different approaches a dataset shall be created containing verdicts from the Dr. Otto Schmidt publisher and freely available judgments from the Internet. As stated above, not every verdict has a guiding principle, thus the summarization of the decision, the tenor, shall also be used to create a bigger dataset. As both have different content, the system will be trained in a multi-task way. The performance impact of this training procedure shall then be compared against the single-task approach.
The goal of this guided research is to create a summarization system directly tailored to German court rulings and evaluate its performance based on standard evaluation metrics [3] and possibly more com-plicated evaluation procedures via Information Extraction. Given an adequate performance level, human evaluation by legal practitioners might also be desirable.

References
[1]    M. Allahyari, S. Pouriyeh, M. Assefi, S. Safaei, E. Trippe, J. Gutierrez, and K. Kochut, "Text Sum­marization Techniques: A Brief Survey," International Journal of Advanced Computer Science and Applications (IJACSA), vol. 8, pp. 397-405, 07 2017.
[2]    0. Klymenko, D. Braun, and F. Matthes, "Automatie Text Summarization: A State-of-the-Art Review," Proceedings of the 22nd International Conference on Enterprise Information Systems (ICEIS 2020), vol. 1, pp. 648-655, 01 2020.
[3]    H. Lin and V. Ng, "Abstractive Summarization: A Survey of the State of the Art," Proceedings of the AAAI Conference on Artificial lntelligence, vol. 33, pp. 9815-9822, 07 2019.
[4]    R. Nallapati, B. Zhou, C. Dos Santos, C. Gulcehre, and B. Xiang, "Abstractive Text Summarization Using Sequence-to-Sequence RNNs and Beyond," Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning, pp. 280-290, 02 2016.
[5]    A. See, P. Liu, and C. Manning, "Get To The Point: Summarization with Pointer-Generator Networks," Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers}, pp. 1073-1083, 01 2017.
[6]    R. B. Soler and J. Kalita, "Abstractive and mixed summarization for long-single documents," ArXiv, 2020.