基于多层最大熵模型的句子主干分析

Skeleton Parsing Based on Multi-layer Maximum Entropy Model

句子主干分析的主要任务是自动识别句子的主干成分。鉴于汉语句子之间成分的相关性,提出一种多层最大熵模型,它的底层最大熵利用句子的上下文特征识别主干词候选项,高层最大熵利用底层最大熵模型的计算结果,结合句子内的远距离特征和句子之间的关系,对底层最大熵模型识别出的主干词候选集进行分析。实验证明,该模型对于简单的主干成分识别正确率较高,对训练语料有一定的依赖;随着语料规模的增长,模型性能缓慢提升。

The main task of Skeleton Parsing is to identify the skeleton of a sentence automatically.Chinese Skeleton Parsing is a key problem in NLP.Because of the interrelation of the skeleton in the same context,a Multi-layer Maximum Entropy Model（MMEM） for the skeleton parsing was proposed.The low-layer ME analyzed skeleton by the context features while the high-layer ME analyzed skeleton by both the result of the low-layer ME and the features between sentences.The experiment showed that MMEM was efficient for Chinese skeleton parsing.A high precision was achieved under a small corpus while it was dependable on the scale of corpus.With the increasing of the corpus,the precision of MMEM improves slowly.