Zero-shot Transfer of Article-aware Legal Outcome Classification for European Court of Human Rights Cases

Findings of the Association for Computational Linguistics: EACL 2023, pages 605–617 May 2-6, 2023 ©2023 Association for Computational Linguistics Zero-shot Transfer of Article-aware Legal Outcome Classification for European Court of Human Rights Cases Santosh T.Y.S.S1, Oana Ichim2, Matthias Grabmair1 1School of Computation, Information, and Technology; Technical University of Munich, Germany 2Graduate Institute of International and Development Studies, Geneva, Switzerland {santosh.tokala, matthias.grabmair}@tum.de oana.ichim@graduateinstitute.ch Abstract In this paper, we cast Legal Judgment Pre- diction on European Court of Human Rights cases into an article-aware classification task, where the case outcome is classified from a combined input of case facts and convention ar- ticles. This configuration facilitates the model learning some legal reasoning ability in map- ping article text to specific case fact text. It also provides an opportunity to evaluate the model’s ability to generalize to zero-shot settings when asked to classify the case outcome with respect to articles not seen during training. We de- vise zero-shot experiments and apply domain adaptation methods based on domain discrim- ination and Wasserstein distance. Our results demonstrate that the article-aware architecture outperforms straightforward fact classification. We also find that domain adaptation methods improve zero-shot transfer performance, with article relatedness and encoder pre-training in- fluencing the effect. 1 Introduction Legal Judgment Prediction (LJP) has recently gained considerable attention in the mainstream NLP community (e.g., Aletras et al. 2016; Chalkidis et al. 2019, 2021, 2022b; Santosh et al. 2022, 2023). In LJP, the outcome of a case should be classified/predicted based on a textual descrip- tion of case facts. In actual legal reasoning, legal practitioners (e.g., advocates, judges) determine rel- evant rules from the sources of law (e.g., statutes, regulations, precedent) that are relevant to the case at hand. They then carry out an analysis to deter- mine which rules apply to the case at hand, and deduce the outcome of the case by applying them. Subsuming case facts under elements of rules given in legal sources plays a critical role in this process. Many current LJP approaches (e.g., Aletras et al. 2016; Chalkidis et al. 2019, 2022b; Santosh et al. 2023) tackle this as a straightforward classification problem with the textual descriptions of case fact as the sole input. This reliance on the model learning statistical correspondences from case fact descrip- tions directly to outcomes neglects the role of legal sources in this relationship. As a consequence, the model may learn sub-optimal fact-outcome pat- terns that are informed by the case distribution in the data rather than learning to align facts with the legal source text containing applicable rules. The models may also attend to outcome-correlating dis- tractors present in the dataset rather than engage in the legal fact-vs-law reasoning that is required of legal practitioners for a proper justification of the outcome (Santosh et al., 2022). This work seeks to remedy this incomplete infer- ence and enable the model to learn more authentic reasoning between rules and case facts by cast- ing LJP into an article-aware classification setting and subjecting it to a zero-shot transfer …

Content truncated to 500 words