Sentence boundary detection in adjudicatory decisions in the United States

We report results of an effort to enable computers to segment US adjudicatory decisions into sentences. We created a data set of 80 court decisions from four different domains. We show that legal decisions are more challenging for existing sentence boundary detection systems than for non-legal texts. Existing sentence boundary detection systems are based on a number of assumptions that do not hold for legal texts, hence their performance is impaired. We show that a general statistical sequence labeling model is capable of learning the definition more efficiently. We have trained a number of conditional random fields models that outperform the traditional sentence boundary detection systems when applied to adjudicatory decisions.