wencolani
wencolani
FB15k的数据是按照每行h \t t \t r 的形式,但是这里load_triple是要求每行为h \t r \t t的形式,将151行relationid = self.__relation2id[line_list[1]] 改为relationid = self.__relation2id[line_list[2]]即可
152行 tailid = self.__entity2id[line_list[2]] 改为 tailid = self.__entity2id[line_list[1]]
在FB15k上的实验结果如下(预测前先对embedding进行norm处理): hit@10(raw): 0.465 hit@10(filter): 0.712 MR(raw): 225.41 MR(filter): 86.304 参数为: dimention=100, margin=1, learning_rate= 0.01, norm = L1, 训练迭代大概300 iteration(或更多,但不超过500) WN18数据集上当时没有测过
Hi, Thanks for your interests. Sorry we didn't try IterE on non sparse version of the dataset. For the code of axiom pool generation, we added it (axiomPools.py). And yes...