is there any negative samples in dataset?

[yzho0907]

` def parse_line(line): line = line.strip().split() sub = line[0] rel = line[1] obj = line[2] val = [1] if len(line) > 3: if line[3] == '-1':

print('entered')

        val = [-1]
return sub, obj, rel, val` in train.txt there is no line list that is over 3, which means no negative samples? i am a bit confused, any idea? thx

[daiquocnguyen]

It is only for the case when a dataset consists of negative samples. You can see such datasets as WN11 and FB13 for the triple classification task.

[yzho0907]

thx, but do u mean there is no any invalid triple in FB15k237 and WN18RR during the training process?

[daiquocnguyen]

During training, invalid triples are sampled by corrupting valid triples, using the common Bernoulli strategy. You can see this in the training protocol section in our paper.

[yzho0907]

is corrupting part in builddata_softplus.py? thx for your kind patience