sapruash / recursivenn Goto Github PK

Since some words' labels may be -1,
in def extract_batch_tree_data from tf_data_utils,
labels_arr.fill(-1) may be not reasonable.(in line 185)
I am appreciated for your reply.

Hello,

In the code, it looks like the files that need to be read are "vocab-cased.txt", "trainsents.txt", "testsents.txt", etc (see below). However, when I download the dataset from http://nlp.stanford.edu/sentiment/index.html (three links on the right hand side), I don't see those files in any of the downloads. Can you kindly point me to where I need to pick the correct set of files? Thanks a ton.

`def load_sentiment_treebank(data_dir,fine_grained):
voc=Vocab(os.path.join(data_dir,'vocab-cased.txt'))

split_paths={}
for split in ['train','test','dev']:
    split_paths[split]=os.path.join(data_dir,split)

fnlist=[tNode.encodetokens,tNode.relabel]
arglist=[voc.encode,fine_grained]
#fnlist,arglist=[tNode.relabel],[fine_grained]

data={}
for split,path in split_paths.iteritems():
    sentencepath=os.path.join(path,'sents.txt')
    treepath=os.path.join(path,'parents.txt')
    labelpath=os.path.join(path,'labels.txt')
    trees=parse_trees(sentencepath,treepath,labelpath)
    if not fine_grained:
        trees=[tree for tree in trees if tree.label != 0]
    trees = [(processTree(tree,fnlist,arglist),tree.label) for tree in trees]
    data[split]=trees

return data,voc`

I get the error about various input data shape ! how is the static computation graph supposed to handle this issue? I taught that I have to use a dynamic computation graph provided in TensorFlow fold and PyTorch for these kinds of tasks. The error I get is this :
(0) Invalid argument: Shapes of all inputs must match: values[0].shape = [17] != values[2].shape = [23] [[node stack (defined at C:\Users\RAKA\Documents\tree_lstm\HSR_tree_lstm\tree_structured_lstm.py:150) ]]
line 150 :
batch_loss = tf.stack(outloss)

Hi, thank you for sharing this code! I wonder if you could point out/explain a bit which part of the code I would need to change to use the SICK dataset? Thank you so much.

Kurt

I just found your code while trying to understand the paper and was wondering if you could help me.
In section 4.2, it is said that the first step in investigating relatedness between two sentences is to produce sentence representations of HL and HR using a Tree-LSTM model over each sentence's parse tree. Afterwards, another neural network is introduced which uses the angle and distance of the obtained sentence representation pairs.
My question would be how exactly are the HL and HR obtained using the Tree-LSTM model?
Section 5.2 also mentions a produced constituency parse using PCFG, how is this used with the Tree-LSTM?

What is the license of the code present in this repository?