Abstract: There are two main deficiencies in the existing methods based on pattern: the amount and reliability of patterns are restricted by the training set, and the methods for scoring patterns and candidate terminologies are too simplified. This paper introduces self-training study mechanism into terminology translation system. This system accomplishes initial study in a pair of terminologies stuff, and selects a more reliable candidate automatically which is retrained in practical running. It adds the model scoring into this system, and improves the method of scoring the candidate terminologies by applying heuristic rules. Experimental results indicate that the improved system has higher performance than the existing systems.

Key words: terminology translation, self-training mechanism, machine learning

摘要： 现有基于模式的术语翻译系统存在2个主要缺点，即学习过程依赖人工标定语料和缺乏对模式的评分以及对候选术语的评分太简单。该文将self-training学习机制引入术语翻译系统，在一对训练语料上完成初始学习，在实际运行中自动选择可靠程度较高的术语重新训练，以改进系统性能。该系统中增加了对模式的评分，利用启发规则，扩充了候选术语的评分方法。实验结果表明，改进后系统的性能高于原有系统。

关键词: 术语翻译, self-training机制, 机器学习

CLC Number: 

• TP312