Ensemble Methods for Word Embedding Model Based on Judicial Text

• Published in: Web Information Systems and Applications Vol. 11817; pp. 309 - 318
• Main Authors: Xia, Chunyu, He, Tieke, Wan, Jiabing, Wang, Hui
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2019; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Ensemble learning; Judicial text; Natural language; Supervised learning; Word embedding
• ISBN: 9783030309510; 3030309517
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-30952-7_31

With the continuous expansion of computer applications, scenarios such as machine translation, speech recognition, and message retrieval depend on the techniques of the natural language processing. As a technique for training word vectors, Word2vec is widely used because it can train word embedding model based on corpus and represent the sentences as vectors according to the training model. However, as an unsupervised learning model, word embedding can only characterize the internal relevance of natural language in non-specific scenarios. For a specific field like judicial, the method of expanding the vector space by creating a professional judicial corpus to enhance the accuracy of similarity calculation is not obvious, and this method is unable to provide further analysis for similarity in cases belonging to the same type. Therefore, based on the original word embedding model, we extract factors such as fines and prison term to help identify the differences, and attach the label of the case to complete supervised ensemble learning. The result of the ensemble model is better than any result of single model in terms of distinguishing whether they are the same type. The experimental result also reveal that the ensemble method can effectively tell the difference between similar cases, and is less sensitive to the details of the training data, the choice of training plan and the contingency of a single inaccurate training run.