HRCSData
收藏资源简介:
该数据集来源于HRCS 2014、2018和2022年度的直接资助项目数据,专注于医学和生物学领域的文本分类任务。数据集经过严格的质量控制,仅包含人工编码的摘要文本,并移除了少于75个字符的摘要记录以确保模型有足够的学习材料。数据划分方面,测试集专门使用2022年的数据以提供现代性能基准。为防止过拟合,实施了基于字符级TF-IDF向量(3-5字元n-gram)和余弦相似度阈值≥0.85的鲁棒去重流程,包括相似性分析、连通分量图聚类和文件顺序优先保留机制。特别强调训练集与测试集的隔离性,通过联合TF-IDF向量空间拟合和跨集相似性比对(≥0.85阈值)来防止数据泄漏,测试集本身仅采用精确标题匹配去重。需注意短小通用标题(如'Studentship')可能因相似度阈值在训练集中被去重。
This dataset is sourced from the direct funding project data of HRCS 2014, 2018 and 2022, focusing on text classification tasks in the medical and biological fields. The dataset underwent strict quality control, only containing manually coded abstract texts, and records with abstracts shorter than 75 characters were removed to ensure sufficient learning materials for models. For data splitting, the test set exclusively uses 2022 data to provide a modern performance benchmark. To prevent overfitting, a robust deduplication process was implemented based on character-level TF-IDF vectors (3-5 character n-grams) and a cosine similarity threshold of ≥0.85, which includes similarity analysis, connected component graph clustering, and a file order-based retention priority mechanism. Particular emphasis is placed on the isolation between the training and test sets: joint TF-IDF vector space fitting and cross-set similarity comparison (with a threshold of ≥0.85) are adopted to prevent data leakage, and the test set itself only uses exact title matching for deduplication. Note that short generic titles (e.g., 'Studentship') may be deduplicated in the training set due to the similarity threshold.



