This builds on the work of Automatic detection of arguments in legal texts; whereas that paper used argumentative texts from multiple domains (including newspapers and social media, despite the title), this work is restricted to the legal domain. Besides detecting argumentative and non-argumentative sentences, premises and conclusions are also detected. Additional features are added to analyze the importance of relations between sentences.
29 admissibility reports and 25 legal cases randomly selected by European Court of Human Rights August 2006 & December 2006. These contain facts, complaints, the law, and final conclusions from judges, expressed in long and complex sentences.
These were manually analyzed by two lawyers to indicate whether they contained arguments. There were 12,904 sentences (10,133 non-argumentative and 2,771 argumentative), which included 2,355 premises and 416 conclusions.
Average accuracy of the maximum entropy model is 82%, using only the information from the current analyzed sentence. (Previous experiments used a naive Bayes model; the increased amount of information in this case meant they could not satisfy the independence assumptions of the naive Bayes classifier). They also experimented with using information in adjacent sentences.
In future work they plan to look at the clause level, instead of the sentence level.