Convolutional-neural-network-based Multilabel Text Classification for Automatic Discrimination of Legal Documents

摘要

Law courts spend too much time reading documents and judging the type of legal cases. This problem becomes more serious as a crime can be classified into several categories at the same time. Thus, legal documents need a multilabel classification. We propose a multilabel text classification model based on multilabel text convolutional neural network (MLTCNN). We scan legal documents and convert them to text data using optical character recognition (OCR) with a charge-coupled device (CCD) sensor. Then, we use Jieba, a word segmentation tool of Chinese letters, and TensorFlow VocabularyProcessor to generate vocabularies. Then, the case description after segmenting each word is mapped into a word index in the vocabularies. We use a word index vector as an input to the MLTCNN. Lastly, we adopt multiple sigmoid functions for multiple binary classifications. The result shows our method to be efficient in finding errors and deviations for similar cases among district courts. This study provides a new method to improve the legal service and to enable fairer law enforcement.