A few-shot transfer learning approach using text-label embedding with legal attributes for law article prediction

摘要

In this paper, we proposed an intelligent law article prediction method to address the data imbalance and missing value problems in law text analysis. The proposed method could predict the various articles of laws that are relevant to a case based on fact description. Typically, the prediction of articles of law has two problems. First, no uniform format exists for law texts. Therefore, some attributes and features may be missing. This problem is termed the missing value problem. Second, certain laws may be referred to infrequently. This problem is called the data imbalance problem. To solve the missing value problem, we applied nonstatic word embedding methods to obtain the weighted vector similarity between words. For addressing the data imbalance problem, we used a weight sharing classification layer to classify the labels according to the relevance of the fact vector to the law article vector of the vector space. We used frequently occurring articles of various laws to train a transfer learning model and shared the weight as the prior knowledge to low-frequency ones to improve classification performance. We compared the functionality of our approach with others for law article prediction. By transferring prior knowledge from frequent cases to rare ones, our method saves legal workers’ time by automatically inferring law articles for cases they seldom deal with. Our experimental results revealed that the proposed article prediction method outperformed the state-of-the-art few-shot article prediction method.