Sightation/SightationRetrieval

SightationRetrieval是一个检索数据集，包含图表图像作为查询，以及相应的描述作为检索文档。数据集包括Top1、Top5、Top10描述和10个难负样本文档。Top1是针对一个图表的4个描述中，根据 sighted 用户评分得分选出的最佳描述。Top5包括4个原始描述加上1个合成描述（修改过的描述，随机缺少一个非最佳句子）。Top10是Top5描述加上5个合成描述。难负样本是从其他图表的Top1描述中选出，根据ALL-MPNET-BASE-V2嵌入空间的余弦相似性排序。

SightationRetrieval is a retrieval dataset consisting of diagram images as queries and corresponding descriptions as retrieval documents. The dataset includes Top1, Top5, Top10 descriptions and 10 hard negative documents. Top1 is the best description selected among 4 descriptions for one diagram based on the rating score annotated by sighted users. Top5 consists of 4 original descriptions plus 1 synthesized description (modified description, missing a random non-best sentence). Top10 is Top5 descriptions plus 5 synthesized descriptions. Hard negatives are selected from the combined pool of Top 1 descriptions for other diagrams, sorted by cosine similarity in the embedding space of ALL-MPNET-BASE-V2.