关键词: 汉语方言辨识, 多示例学习, 多样性密度, k近邻, 平均最近距离

Abstract: In order to solve the problem that designing Chinese dialect model singly and improve the performance of dialect identification, an approach of Chinese dialect identification based on combination diverse density is presented. Diverse density is a classical algorithm of multi-instance learning. Combination diverse density is a improved application algorithm based on it. The new method firstly pre-classify one kind dialect into several little classes. Secondly generate every little class into multi-instance bags. Then use EM-DD for multi-instance learning and get various diverse density points as a dialect model. Finally put forward average recent distance algorithm for classification. The method can get a complete and full model in training part, and consider the influence of every instance in unseen bags in pattern classification part. Finally the efficiency of the system is improved.

Key words: Chinese dialect identification, multi-instance learning, diverse density, k-means, average recent distance