| Summary: | 碩士 === 國立臺北大學 === 資訊工程學系 === 107 === This thesis proposes a law article prediction approach in legal intelligent, which solves the law articles
imbalance problem and the missing value problem of the judgment. The proposed approach predicted the
involved law articles by given the fact description of the case. However, the factual description of the judgment
has been incomplete after pre-process because the format is no uniform, and the legal data set in real cases
often faces the situation of imbalanced data, which means some law articles have been less involved in the
case and cause the worse performance when predicts the low-frequency law articles. To address these problem,
this thesis applies the law article description as the label attribute, based on the property of the vector space to
get over the missing value problem by learning a representative embedding vector through training the nonstatic
word embedding by the vector similarity weighted mechanism. For the imbalance problem, we use a
weight sharing classification layer which classifies the label according to the relevance between the fact vector
and law article vector of the vector space. We also used the transfer learning to train the model by the highfrequency
law articles first, then sharing the weight as the prior knowledge to the low-frequency one to
improve the classification performance. The proposed approach outperforms the performance on few-shot law
article prediction.
|
|---|
