Few-Shot Learning for Chinese Legal Controversial Issues Classification

Abstract

Chinese courts organize debates surrounding controversial issues along with the gradual formation of the new procedural system. With the progress of China's judicial reform, more than 80 million judgement documents have been made public online. Similar controversial issues identified in and among the massive public judgment documents are of significant value for judges in their trial work. Hence, homogeneous controversial issues classification becomes the basis for similar cases retrieval. However, controversial issues follow the power-law distribution, not all of them are within the labels provided by manual annotation and their categories cannot be exhausted. In order to generalize those unfamiliar categories without necessitating extensive retraining, we propose a controversial issues classification algorithm based on few-shot learning. Two few-shot learning algorithms are proposed for our controversial issues problem, Relation Network and Induction Network, respectively. With only a handful of given instances, both of them have shown excellent results on the two datasets, which proves their effectiveness in adapting to accommodating new categories not seen in training. The proposed method provides trial assistance for judges, promotes the dissemination of experience and improves fairness of adjudication. INDEX TERMS Controversial issues, few-shot learning, text classification, power-law, BERT.

A Survey of Research on Data Analytics-Based Legal Tech

“…Examples of document type target variables are case type, complaint type, accusation type, and topic type, etc. [57,62,72,77,78].…”

Hierarchically Structured Network With Attention Convolution and Embedding Propagation for Imbalanced Few-Shot Learning

“…= ( − ) −1 (5) Where a∈R is a controlling variable, and I is the identity matrix. Next, we compute embeddings by (6). =…”

Messier than Oil: Assessing Data Advantage in Military AI

“…Beyond foundational research, significant practical strides in few-shot learning and synthetic data have yet to be applied (and exist neither within the U.S. or Chinese government to the best of the authors' knowledge), despite a demonstrated interest. 53 There is still time before both sides start making effective use of them. If proven practical and applicable for military needs, these approaches would enable development of AI without reliance on massive datasets and might mean, in effect, that data is not the new oil; rather, it is no more than yesterday's whale oil.…”