基于语句压缩的中文语义依存分析

Chinese semantic dependency parsing based on sentence compression

针对当前中文语义依存分析中耗时长、准确率低的问题,提出了一种基于语句压缩进行中文语义依存分析的方法。在此方案中,首先通过开源工具CRF++训练得到特定的序列化标签压缩模型,通过此模型得到任意输入句子的主干信息,并为下一步提供候选集;其次,选取原句和压缩后句子中的词性、上下文等特征,使用条件随机场对其中的语义依存关系进行识别;最后进行谓语消歧和句子回溯。实验以Co NLL 2009公开任务中的公有语料作为数据集,与传统的直接使用基于图的语义依存分析方法相比,本方案的处理时间缩短了80%,精确率提高了3.48%,综合指标提高了2.11%。

To solve the problem that the method currently used to parse Chinese semantic dependency is time-consuming and inaccurate, a solution based on sentence compression was implemented. Firstly, a special sequence labeling compression model was created by open source tool CRF ＋ ＋ with training data, which contributed to detect the backbone information of any sentences, and provided the candidate for the next step. Secondly, the features of original sentence and compressed sentence such as characteristic and context were selected, the semantic dependency relation was recognized by using CRF（ Conditional Random Field）. Finally, predicate sense disambiguation and sentence recovery were implemented. The data set was based on Co NLL 2009 shared task. Compared with the traditional semantic dependency parsing method based on graph, processing time is reduced by 80%, the precision rate gets a 3. 48% improvement, the F-measure gets a 2. 11% improvement.