A Few-Shot Transfer Learning Approach Using Text-label Embedding with Legal Attributes for Law Article Prediction

EasyChair Preprint 1344

Abstract

Law article prediction is to determine the appropriate law article according to the fact descriptions, which is useful for legal auxiliary system when given a fact description of the case. A problem of legal charge prediction is often faced with the imbalance data problem, which represent the few-shot category with limited case. To address the imbalance data, this paper introduces law article as label embedding to improve the relevance between fact and law articles. More specifically, this paper applies the weight sharing mechanism of transfer learning to utilize the data with high frequency to model the projection between fact and law articles, also as a prior knowledge to achieve law article classification for case with low frequency data. This paper employs an attention mechanism based on text similarity to produce the fact context vector, and then infer the law article associated with the case by utilize a label set independent projection layer. The experimental results show that, the label embedding of law article can improve the prediction performance, and transfer the weight of projection layer to few-shot data can achieves better performance on few-shot data classification.

Keyphrases: Natural Language Processing, deep learning, few-shot learning, law article prediction, legal judgment prediction

@booklet{EasyChair:1344,
  author    = {Yuh-Shyan Chen and Shin-Wei Chiang and Tong-Ying Juang},
  title     = {A Few-Shot Transfer Learning Approach Using Text-label Embedding with Legal Attributes for Law Article Prediction},
  howpublished = {EasyChair Preprint 1344},
  year      = {EasyChair, 2019}}