面向序列迁移学习的似然比模型选择方法

Model selection with likelihood ratio for sequence transfer learning

为了解决迁移学习的欠适配问题,将粒模型作为候选模型的集合,通过模型选择的方式引入目标域的辅助模型中包含的标注规则,提出粒模型推断中基于似然比的模型选择方法(likelihood ratio model selection,LRMS),实现了辅助模型与粒模型的融合。LRMS保持基于Viterbi算法的标注模型对整条序列进行计算的模式,避免了候选标注器对上下文关系的破坏。通过大量词性标注实验表明LRMS在每个迁移学习任务中都有准确率的提高,从而证明似然比模型选择是一种有效的解决欠适配问题的方法。

To solve the under-adaptation problem of transfer learning,in this paper the granular model is used as a set of candidate models,and labeling rules contained in minor for target domain models is introduced by a model selection method. We propose a Likelihood Ratio based Model Selection method（LRMS） for the inference of granular model,which implements the fusion of minor models with the granular model. LRMS keeps the single-path calculating of Viterbibased sequence labeling model,which avoid the violation of contextual connections. In empirical experiments on part-of-speech tagging,LRMS improves the accuracy in every transfer learning task,therefore,the effectiveness of LRMS in solving the under-adaptation problem is verified.