基于语义图优化算法的中文微博观点摘要研究

为从海量微博中高效地获取不同话题下的关键信息,微博观点摘要成为自然语言处理领域近期研究的热点之一。基线方法基于TF-IDF算法抽取微博句中的关键词,并据此计算微博的重要性分数,直接筛选出观点摘要;朴素改进方法在基线方法的基础上,增加了情感分类步骤,并利用微博句之间的语义距离,将摘要句候选集中语义重复、重要度较小的句子去除,生成观点摘要;基于语义图优化算法的方法在朴素改进方法的基础上,利用微博句的重要性分数及微博句之间的语义距离构建语义图结构,并通过图优化算法筛选出观点摘要。朴素改进方法在COAE2016评测任务一测试数据集上,10个话题的平均ROUGE-1值达到26.39%,平均ROUGE-2值达到0.68%,平均ROUGE-SU4值达到5.69%,且评测官方公布结果显示,该方法在9项评价指标中获得6项最佳性能。基于语义图优化算法的方法在评测样例数据集上进行了实验,结果显示,该方法比朴素改进方法在ROUGE-1,ROUGE-2,ROUGE-SU4值上分别提升了0.63%, 1.51%, 2.69%。

Semantic graph optimization algorithm based chinesemicroblog opinion summarization

To obtain key information in different topics efficiently, microblog opinion summarization has been a hot spot in natural language processing recently. The baseline method of this paper extracts keywordsusing TF-IDF algorithm, and calculate the importance scores of microblogs to filter out opinion summarization directly; the naive improved methodadded a step of sentiment classification, andremove microblogs which are of low importance and high semantic repetitionusing semantic distance between microblogs to generate opinion summarization;the method based on semantic graph optimization algorithm constructs a complete graph using importance scores and semantic distance of microblogs, and filters out the opinion summarization using graph optimization algorithm. According to the official result of evaluation,on the test dataset of COAE2016, the average ROUGE-1 value, ROUGE-2 value and ROUGE-SU4 value of 10topics using the naive improved methodreached 26.39%, 0.68% and 5.69% respectively, and got 6 max values out of 9 kinds of evaluation index. Besides, the results of experiments done on COAE2016 sample datasetshows that by using the method based on semantic graph optimization algorithmthe ROUGE-1 value, ROUGE-2 value and ROUGE-SU4 value increased by 0.63%, 1.51%, 2.69% respectively.