Datasets:
license: apache-2.0
task_categories:
- summarization
language:
- zh
size_categories:
- 100K<n<1M
一个来自K-SportsSum：https://github.com/krystalan/k-sportssum 的实现，原作者给出了思路，但并未实现其具体过程，此数据集是对该数据集“新闻与评论句子根据相似度搭配”部分的实现。 方法是：遍历新闻句子，以类似指针的方式获取新闻句子的时间信息（如果有的话），然后将两个指针作为范围，将范围内的新闻句子遍历，在同一时间范围之内查找评论句子，评分后选择最高的结果，并删除该句以防止重复，最终获得一句新闻搭配一句评论的结果。 我使用了bert—Score和ROUGE指标，按照7:3加权计算分数。 建议 数据集内给出了该搭配的指标，请考虑使用平均数等方式过滤掉较低的坏搭配。 An implementation from K-SportsSum: https://github.com/krystalan/k-sportssum was used to implement the "news and comment sentences paired based on similarity" section of the dataset. The original author provided the idea but did not implement the specific process. The method involves traversing the news sentences and obtaining the time information of the news sentences (if any) in a pointer-like way. Then, two pointers are used as the range to traverse the news sentences. Within the same time range, the comment sentences are searched, and the highest scored result is selected and deleted to avoid duplication, resulting in one news sentence paired with one comment sentence. BERT-Score and ROUGE metrics were used with a 7:3 weighted score calculation. It is recommended to use averaging or other methods to filter out poor pairings based on the metrics provided within the dataset.