fanyangxyz/neural-lp

Differentiable Learning of Logical Rules for Knowledge Base Reasoning

neural-lp's Issues

Why we need both query and M?

We want to predict the probability over all the M.
What is the function of query as place_holder input.
Thank you.
@fanyangxyz

Small edits to instructions mentioned

Hi

Thank you for this code.

I think these edits would help people running the code:

.eval/collect_all_facts.sh datasets/family
python eval/get_truths.py datasets/family
python eval/evaluate.py --preds=exps/family/test_predictions.py --truths=datasets/family/truths.pckl

u_{T+1}: Final step in the recurrent formulation

@fanyangxyz : In the code starting at:

Neural-LP/src/model.py

Line 148 in cbb6a9d

for t in xrange(self.num_step):

, I do not see the final step in the recurrent formulation, namely, Eq 8 in your NIPS paper.

I am a little confused in this part. Could you please explain ?

Further, I assume https://github.com/fanyangxyz/Neural-LP/blob/master/src/model.py#L149 is Eq 11 in the paper. However, the indices seem to not match the equation. In the equation, you have softmax([h_0 .. h_{t-1}]^T h_t) but the code seems to suggest softmax([h_0 .. h_{t}]^T h_t)

It would be really great if you could explain this discrepancy. Thanks

Could you please teach me what M_R_k means?

M_R_k is in equation 2:

I know M_R is a matrix of shape |E|×|E| as the paper write:

Thank you.
Thank you.
Thank you.
@fanyangxyz

Do you think Neural-LP can solve this kind of question?

1,
A lives in C. (A, liveInCity, C)
B lives in C. (B, liveInCity, C)
Does A and B live in the same city?

2,
A is 180 cm height. (A, height, 180)
B is 175 cm height. (B, height, 175)
Is B taller than A?

Thank you very much.@fanyangxyz

N-ary relation logic rule?

How can it be used to n-ary relation graphs and generate the logic rules

Could you please teach me what the alpha and beta of the paper correspond to the code?

What are α and β of the paper in the code?
I guess they may be in src/model.py

Thank you.
Thank you.
Thank you.
@fanyangxyz

Inductive Datasets ?

Hi,
Nice work!
Can I know where the inductive dataset is located? Or the exact parameters (random sample of test entities etc ) required to build the inductive dataset as mentioned in the paper ?

Thanks

Other datasets / experimental framework

@fanyangxyz Do you plan to release the other datasets and the experimental framework of the NIPS 2017 paper ? It would be very helpful for research purposes.

Thanks!

Negative facts coverage

Hi @fanyangxyz ,

Can Neural-LP be modelled to generate the set of rules in the presence of incorrect and negative facts in the database? Basically, the rules should cover as many facts as possible from the positive (facts.txt) and as little negatives as possible from a negatives.txt

Running problem

I tried to run the code as the readme suggested. I got the following error:

Traceback (most recent call last):
File "src/main.py", line 6, in
from model import Learner
File "~/Neural-LP-master/src/model.py", line 207
lambda (grad, var): self._clip_if_not_None(grad, var, -5., 5.), gvs)
^
SyntaxError: invalid syntax

What is the difference between train txt and facts txt?

Thank you
@fanyangxyz

question about formula (2) and (5)

Hi @fanyangxyz ,

I am confused why formula (2) can be converted into to formula (5). It seems like these two formulae are not equivalent and there is no explanation in paper. Would you mind tell me more details about this?

Question answering task

I was wondering how you performed QA task explained in Section 4.4 of your paper.
I could process the dataset to match the input format of Neural LP and I could produce rules for the dataset in the exps/demo/directory, but I do not know what to do next. Could you please elaborate more on this task?

Why matrix multiplications of M can lead to the logical rule inference?

Could not understand these lines of code, Thank you..

https://github.com/fanyangxyz/Neural-LP/blob/master/src/model.py#L213-L214

feed[self.queries] = [[q] * (self.num_step-1) + [self.num_query] for q in qq]

Why use many copy of q?
Thank you @fanyangxyz

Grid path finding

which dataset was used in Section 4.2 "Grid path finding"? I can't find the code of this task in your dir, could u give some more details？ thank u~

Question about table 6

I want to know the fairness of this experiment.
The article describes how to handle test sets and training sets. But Neural-LP reasons based on known facts and rules. The facts is splited from the initial train data and contains tuples share entities with selected test tuples.
In order to ensure fairness, different models should use the same data. I want to know the data you use to train TransE because I think it is not the union of facts and train for Neural-LP. What's your interpretation of the fairness of this experiment?