基于结构化学习的语句压缩研究

近年来随着各类信息的日益增多,语句压缩作为自动摘要的重要部分也越来越引起研究者的关注。然而当前针对语句压缩的研究才刚刚展开,存在压缩效果不佳、没有统一的自动评测指标等问题。该文在简单的删除单词的方法框架下,采用基于特征权重的最大边缘训练的结构化学习方法实现语句压缩。同时该文还提出了两种新的自动评价指标(N-Gram和BLEU)来评价语句压缩的性能。实验结果表明,采用结构化学习方法能够在保持较好压缩率的情况下保留源语句的主要信息,并且新提出的两个评价指标能够有效反映语句压缩性能。

Sentence Compression Based on Structured Learning

With the rapid growth of information in recent years, sentence compression as a subtask of summarization attracts more attention. However, the research on sentence compression is in its initial stagethe performance is still beyond satisfaction and it suffers from unavailability of uniformed evaluation metrics. This paper falls in the framework of simply shortening a sentence by deleting words or constituents, and adopts structured learning approach coupled with the large margin training process. Further more, it proposes two new automatic evaluation metrics (N-Gram and BLEU) for sentence compression. Experimental results show that using of structured learning have maintained a good compression ratio while reserving the main information of source sentence. It also shows that the proposed two evaluation metrics effectively reflect the quality of sentence compression.
Key wordssentence compression;structured learning;automatic evaluation