基于栈式降噪自编码和词嵌入表示的维吾尔语零指代消解

Zero Pronoun Resolution of Uyghur Based on Stacked Denoising Autoencoder and Word Embedding

针对维吾尔语零指代现象,提出采用栈式降噪自编码的深度学习机制进行维吾尔语零指代消解。首先由大规模无标注维吾尔语语料训练得到富含语义和句法信息的词嵌入表示,将其作为候选先行语和缺省零代词的语义特征;其次根据维吾尔语语言特点,抽取14项针对零指代消解任务的手工设计特征;然后融合word embedding特征和14项hand-crafted特征作为栈式降噪自编码的输入,最后经过无监督逐层贪婪的预训练和有监督的微调过程,使用softmax进行分类完成维吾尔语零指代消解任务。实验结果表明,与传统栈式自编码、浅层机器学习的支持向量机和人工神经网络相比,栈式降噪自编码的F值分别提高了4.450%、10.032%和8.140%,实验结果验证了该方法的有效性及栈式降噪自编码在任务中具备挖掘高层面鲁棒性语义特征的优势。

Adopting deep learning mechanism,this paper apply Stacked Denoising Autoencoder（SDAE）to deal with Uyghur zero pronoun anaphora phenomenon.Firstly,word embedding trained on large-scale unlabeled Uyghur corpus is used as semantic features of candidate antecedents and zero pronouns.Secondly,according to Uyghur characteristics,we extract 14 hand-crafted features for zero pronoun resolution.Experimental results show that,compared to SAE（Stacked Autoencoder）,SVM and ANN,the F value of SDAE is increased by 4.450%,10.032% and 8.140%,respectively.