NELL-995是卡内基梅隆大学开发的知识图谱数据集，包含超过百万实体和数十万关系的事实陈述，用于训练机器学习模型理解与生成知识图谱。
The dataset format is as follows:
- **raw.kb**: Contains the raw KB data from the NELL system.
- **kb_env_rl.txt**: Includes inverse triples of all triples in raw.kb. This file serves as the KG for reasoning purposes.
- **entity2vec.bern/relation2vec.bern**: TransE embeddings used to represent RL states, which can be trained using TransX implementations by thunlpt (though specific implementation details are not provided here).
The tasks directory contains specific reasoning relations and their associated files:
- **tasks/${relation}/*.vec**: Trained TransH Embeddings.
- **tasks/${relation}/*.vec_D**: Trained TransD Embeddings.
- **tasks/${relation}/*.bern**: Trained TransR Embeddings, trained using the specified KB embeddings.
- **tasks/${relation}/*.unif**: Trained TransE Embeddings.
Additionally, there are directories for each relation containing:
- **transX**: Triples used to train the KB embedding models.
- **train.pairs** and **test.pairs**: Training and test triples in PRA format respectively.
- **path_to_use.txt**: Reasoning paths discovered by the RL agent.
- **path_stats.txt**: Path frequency of randomised BFS.
5星
