Automatic Summarization of Legal Bills: A Comparative Analysis of Classical Extractive Approaches

A Legal Document is typically very long and structurally rich, which makes a quick understanding of the document very difficult. One way to deal with this difficulty is to manually summarize these documents with the help of legal experts. However, this is a costly and time consuming process. Design of automatic summarization approaches could be the key, to achieve a more time and cost effective solution to this problem. In this work, a detailed comparative analysis of multiple classical extractive summarization techniques is presented, with respect to Recall-Oriented Understudy for Gisting Evaluation metrics (ROUGE), Bilingual evaluation understudy metrics (BLEU) and Cosine Similarity scores, which are widely used metrics for evaluating automatic summarization approaches. The experimental analysis is performed on a publicly available benchmark dataset. From the experimental results, it has been observed that graph based summarization techniques perform well in general, across all the evaluation metrics. Another important observation is that, in addition to word frequency, considering other key contextual information can boost the performance of automatic summarization techniques. This comparative analysis work can serve as a baseline on the benchmark legal dataset, which is expected to be helpful for further research in this domain.