Despite their growing presence in everyday life, contracts remain notoriously inaccessible to laymen. What for? Here, a corpus analysis (n 10 million words) revealed that contracts contain surprisingly high proportions of some hard-to-deal features – including low-frequency jargon, middle-embedded clauses (which lead to long-distance syntactic dependencies), passive linguistic structures, and non-standard capital letters – compared to nine other basic genres of written and spoken English. Two experiments (N = 184) further showed that extracts containing these characteristics were retrieved and understood at lower rates than extracts without these characteristics, even for experienced readers, and that clauses embedded in the center inhibited memory more than other characteristics. These findings: (a) undermine the specialized concepts of legal theory that law is a system based on a thorough knowledge of technical concepts; (b) indicate that these processing difficulties are largely due to working memory limitations imposed by syntax dependencies over long distances (i.e. poor writing), as opposed to a simple lack of legal skills; and (c) proposing the removal of problematic features from legal texts would be cumbersome and beneficial to society as a whole. Badges are live and dynamically updated with the latest ranking in this document. Laws and their interpretations, legal arguments and agreements are usually expressed in writing, which leads to the creation of extensive bodies of legal texts. Their analysis, which is at the heart of legal practice, becomes more sophisticated as these collections grow. Natural language understanding (NLU) technologies can be a valuable tool to assist lawyers in these efforts.
However, their usefulness depends largely on the possibility of generalizing current models of the state of the art to different tasks in the legal field. To answer this currently open question, we present the LexGLUE benchmark (Legal General Language Understanding Evaluation), a collection of datasets for evaluating in a standardized way the performance of the model in various NLU legal tasks. We also provide an assessment and analysis of several generic and rights-based models, showing that they consistently improve performance across multiple tasks. Individuals and organizations working with arXivLabs have embraced and embraced our values of openness, community, excellence, and user privacy. arXiv is committed to these values and only works with partners who adhere to them. arXivLabs is a framework that allows employees to develop and share new arXiv features directly on our website. Paste the marker at the top of your GitHub README.md file to illustrate the model`s performance. Upload an image to customize your repository`s social media preview. Images must be at least 640×320px (1280×640px for the best view).