search

shehzaadzd / MINERVA

Meandering In Networks of Entities to Reach Verisimilar Answers
Apache License 2.0
312 stars 88 forks source link

Encode the relation and entity ? #22

Closed ProQianXiao closed 5 years ago

ProQianXiao commented 5 years ago

Hello, I read your paper and codes and there is a confusion.

About encoding relation and entity, here is my understanding. You don't encode them as vectors, instead using embedding matrices r and e and looking up in the embedding matrices according to their ids. And tf.nn.embedding_lookup() function could train the parameters in the embedding matrices.

Is that right?

shehzaadzd commented 5 years ago

I'm not sure I understand your question. How is this different from encoding e and r as vectors?

ProQianXiao commented 5 years ago

Sorry, maybe I didn't describe my question clearly. My understangding is that (taking encoding relation as an example): you initialize a relation matrice r randomly like this: default and when you encode relation (such as "LocatedIn" in the country dataset), you will find the corresponding id in the "relation_vocab.json" file, assume its id is 2. Then using tf.nn.embedding_lookup(r,2) will get the third line in the matrice, and that is the encode of the relation "LocatedIn".

I said that "You don't encode them as vectors" means that you didn't encode them using NLP, without considering their semantic information.

Is that right?

shehzaadzd commented 5 years ago

If by NLP you mean we didn't use sentences "Boston is located in the US" to encode the vector for LocatedIn, then you're correct. We created relation embeddings from random and trained the embedding parameters using the RL objective. I hope that helps! :)

ProQianXiao commented 5 years ago

Thanks very much, that helps.