基于矩阵分解和子模最大化的微博新闻摘要方法
发布时间：2018-09-12 19:51
【摘要】：针对面向微博的中文新闻摘要的主要挑战,提出了一种将矩阵分解与子模最大化相结合的新闻自动摘要方法。该方法首先利用正交矩阵分解模型得到新闻文本潜语义向量,解决了短文本信息稀疏问题,并使投影方向近似正交以减少冗余;然后从相关性和多样性等方面评估新闻语句集合,该评估函数由多个单调子模函数和一个评估语句不相似度的非子模函数组成;最后设计贪心算法生成最终摘要。在NLPCC2015数据集上的实验结果表明,该方法能有效提高面向微博的新闻自动摘要质量,ROUGE得分超过其他基线系统。
[Abstract]:Aiming at the main challenge of Weibo's Chinese news abstract, this paper proposes an automatic news digest method which combines matrix decomposition with submodule maximization. Firstly, the latent semantic vector of news text is obtained by using orthogonal matrix decomposition model, which solves the problem of sparse information in short text, and makes the projection direction approximate orthogonal to reduce redundancy. Then the set of news statements is evaluated from the aspects of correlation and diversity. The evaluation function is composed of several monotonic submodules and a non-submodule function to evaluate the dissimilarity of statements. Finally, a greedy algorithm is designed to generate the final summary. The experimental results on the NLPCC2015 dataset show that this method can effectively improve the quality of automatic news abstracts for Weibo and the score of group is higher than that of other baseline systems.
【作者单位】： 武汉大学计算机学院;
【基金】：国家社科重大招标计划资助项目(11&ZD189) 国家自然科学基金面上资助项目(61373108)
【分类号】：TP391.1
本文编号：2240056
[Abstract]:Aiming at the main challenge of Weibo's Chinese news abstract, this paper proposes an automatic news digest method which combines matrix decomposition with submodule maximization. Firstly, the latent semantic vector of news text is obtained by using orthogonal matrix decomposition model, which solves the problem of sparse information in short text, and makes the projection direction approximate orthogonal to reduce redundancy. Then the set of news statements is evaluated from the aspects of correlation and diversity. The evaluation function is composed of several monotonic submodules and a non-submodule function to evaluate the dissimilarity of statements. Finally, a greedy algorithm is designed to generate the final summary. The experimental results on the NLPCC2015 dataset show that this method can effectively improve the quality of automatic news abstracts for Weibo and the score of group is higher than that of other baseline systems.
【作者单位】： 武汉大学计算机学院;
【基金】：国家社科重大招标计划资助项目(11&ZD189) 国家自然科学基金面上资助项目(61373108)
【分类号】：TP391.1
【相似文献】
相关硕士学位论文 前1条
1 高亚奇;基于判别特征回归的子模优化跟踪算法[D];大连理工大学;2016年
，本文编号：2240056
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2240056.html
