An Approach of Rhetorical Status Recognition for Judgments in Court Documents using Deep Learning Models

In a court document, the rhetorical status of a sentence conveys the intention of the sentence, whether is is a claim or contains supporting evidences, thus, is beneficial to court document processing systems, for example, court document retrieval systems. Besides, rhetorical structure analysis has high-impact applications in natural language processing, for instances, text summarization, sentiment analysis, question answering. The output structures of the analysis contain high-level relationship between clauses and so provides valuable information. Despite of a wide range of applications and the necessity for automatic court document processing, automatic rhetorical structure analysis has not been well noticed in the legal domain. We propose to use deep learning models for tackling the task of recognizing the rhetorical status of each sentence in a court document. Deep learning has been shown effective towards natural language processing tasks including discourse analysis. We have achieved promising results for the task, which suggests the applicability of artificial neural module embedding rhetorical information for other tasks, for example, summarization and information retrieval.