Abstract: This paper proposes a new question-focused summarization method based on Semantic Relational Triple(SRT). After syntax parsing, it abstracts the semantic relational triple both from sentences and questions, and calculates the similarity of triples by using search engine. The similarities of sentences and questions are calculated to form the final question-focused summarization. Experiments on authoritative corpus show that this feature outperforms classical features and its ROUGE-4 score rises by 46.4%. The system using the single feature ranks 10/32 in a global evaluation.

Key words: Semantic Relational Triple(SRT), question-focused summarization, automatic textual summarization

摘要： 提出一种新的基于语义关系的特征，在句法分析的基础上，抽取句子及问题的语义关系三元组，利用搜索引擎计算三元组的相似度，在此基础上计算得到句子和问题的相似度，抽取句子形成问答式文摘。权威语料上的实验证明，使用该特征在各项文摘性能指标上超越了经典的TF*IDF方法，ROUGE-4指标提高了46.4%，而且由该特征单独编制的系统，在32家单位参加的国际评测中ROUGE-L指标排名为第10。

关键词: 语义关系三元组, 问答式摘要, 文本自动摘要

CLC Number: 

• TP391