Recursive Top-Down Production for Sentence Generation with Latent Trees

Code for the paper: http://arxiv.org/abs/2010.04704

To run:

python scan.py 

runs the code on the simple split of SCAN.

To run on the other splits, replace <split> with add_jump, add_turn_left, or length

python scan.py \
    --src_path_train data/$SPLIT/train_wo_valid_random.src \
    --trg_path_train data/$SPLIT/train_wo_valid_random.trg \
    --src_path_valid data/$SPLIT/valid.random.src \
    --trg_path_valid data/$SPLIT/valid.random.trg \
    --src_path_test  data/$SPLIT/test.src \
    --trg_path_test  data/$SPLIT/test.trg

This is where the algorithm described in the paper is implemented in the function forward_ctreec. One particular implementation detail is the order of the nodes in the array log_probs for Loss. The order of the flattened word emission probabilities from the tree are in in-order traversal:

ctreec_test.py mocks a tree as shown below, and marginalises over trees of length 4.

              7
        /           \
       3             11       
    /     \        /     \    
   1       5      9       13   
  / \     / \    / \     /   \  
 0   2   4   6  8   10  14   15

Lines 77-81 in the file are the indices that the probabilities need to be extracted from to sum over the correct trees.

For example, on line 80, a tree represented by [3, 8, 10, 13] would be the following:

             7
       /           \
      3             11       
                 /     \    
                9       13   
               / \    
              8   10