【作者】 韩冰；

【导师】 陈惠鹏；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2015， 硕士

【摘要】 文本提出了一种基于感知器的中文分词增量训练方法,可在训练好的模型基础上添加目标领域标注数据继续训练,解决了大规模切分数据难于共享,源领域与目标领域数据混合需要重新训练等问题。在CTB5.0新闻领域和诛仙小说领域的实验表明,增量训练可以有效解决领域适应问题。达到与传统数据混合相类似的效果。进一步的实验分析表明,增量训练产生的模型远小于混合训练方式产生的模型,增量训练在小规模数据集上能够更快速的完成训练。针对基于感知器的中文分词模型过大问题,本文实现了一种启发式的特征筛选方法,以迭代训练中参数的更新次数为阈值进行特征筛选。这种方式有效的避免了自然语言处理中的长尾效应,能够更好的选择有用特征。在基于感知器的中文分词、词性标注、依存句法分析三个任务上的实验表明,在准确度几乎不变的前提下,均可以使模型成倍的缩小。基于增量训练算法和模型压缩算法,本文实现了一个在线的个性化分词服务系统。用户可以上传目标领域词典和训练语料,系统在后台自动完成增量训练算法,为用户生成目标领域增量模型,提供个性化分词服务。本文提出的增量训练方式保证了多个增量模型之间互不影响,同时有效控制了训练时间和模型大小。更多还原

【Abstract】 In this paper, we propose an incremental-styled learning scheme in perceptron based Chinese word segmentation. Our method can perform continuous training over a fine tuned source domain model. Such scheme allows deliver ing model without annotated data and without re-training on these data. The experimental results on CTB5.0 and Zhuxian shows the scheme we proposed can significantly improve adaptation performance on Chinese word segmentation and achieve comparable performance with traditional method. And our further experimental analysis shows that our method can significantly reduce the resulted model size and obtain segmentation model with less time consumption.The model size of perceptron based Chinese word segmentation is usually too large. To solve this, we achieve a heuristic method to filter features, which filters upon the times of the parameter during the training phase. This method can effectively avoid the long-tail effect in NLP, which can help choosing relevant features better. The experimental results on perceptron based Chinese word segmentation and part of speech and dependency parsing shows the method we proposed can efficiently compress the model size with little accuracy reduction.We build an online custmoized Chinese word segmentation system based on the incremental-styled learning scheme and the model compression algorithm. Users can upload the target domain dictionary and annotated data. Our system will execute the incremental-styled learning algorithm in the background to train the incremental-styled target domain model for users. And then our system can provide customized service for users. Benefitting from the incremental-styled learning algorithm and model compression algorithm, multiple incremental-styled models are independent of each other. And the model size and training time is also well guaranteed.更多还原