Understanding train-roto-ptrs.txt file

[wingedRuslan]

Hi Ratish,

Thanks a lot for the insightful research paper as well as making the codebase publicly available! I was able to train the model on boxscore-data and then used it for inference. Now I am interested in training your model on my dataset and then use it for text generation.

Unfortunately, I encountered a problem with the following step (from the README page):

The train-roto-ptrs.txt file is available along with the dataset and can also be created by the following command:

python data_utils.py -mode ptrs -input_path $BASE/rotowire/train.json -train_content_plan $BASE/rotowire/inter/train_content_plan.txt -output_fi $BASE/rotowire/train-roto-ptrs.txt

Since my dataset is different from boxscore-data, I basically have to perform transformations manually to my dataset in order to have it in a suitable format for model training.

1. "The input dataset for data2text-plan-py can be created by running the script create_dataset.py in scripts folder."

I have successfully prepared my dataset in the same format as files in boxscore-data/rotowire/, namely, the following files: src_train.txt, train_content_plan.txt, tgt_train.txt, inter/train_content_plan.txt and src_valid.txt, tgt_valid.txt, valid_content_plan.txt, inter/valid_content_plan.txt and test/src_test.txt , test/tgt_test.txt. The structure of these files is the same as in files obtained from boxscore-data

2. [Preprocessing] I could not run preprocess.py script, since to run this script I need train-roto-ptrs.txt file.

The function for creating this file is:

data2text-plan-py/data_utils.py

Line 574 in 4b74535

def make_pointerfi(outfi, inp_file="rotowire/train.json", content_plan_inp="inter/train_content_plan", resolve_prons=False):

Because I have another dataset, I can not use the above-mentioned function, thus I need to create train-roto-ptrs.txt file by myself. Unfortunately, going multiple times through the function implementation and analyzing the content of a file (please see comments), I could not figure out how to create such file from my dataset.

Can you please elaborate on the purposes of the train-roto-ptrs.txt file and briefly describe the steps on how it was created?

This is my current bottleneck and I would highly appreciate your help with this issue!

Many thanks in advance,
Ruslan

[wingedRuslan]

Analyzing the content of a file

Looking into the content of the original train-roto-ptrs.txt file did not help to understand how the file was created.

From my perspective they are not related to any other information from other files:

@ratishsp, Could you please help me to figure out how the train-roto-ptrs.txt file was created?

[ratishsp]

Hi Ruslan,
Glad to know that you find the paper and code useful!
The logic for train-roto-ptrs.txt is in method

data2text-plan-py/data_utils.py

Line 574 in 4b74535

def make_pointerfi(outfi, inp_file="rotowire/train.json", content_plan_inp="inter/train_content_plan", resolve_prons=False):

The core idea is to provide supervision while training the copy mechanism.

data2text-plan-py/onmt/modules/CopyGenerator.py

Line 199 in 4b74535

switch_loss = self.switch_loss_criterion(p_copy, align.ne(0).float().view(-1, 1))

The entries in the train_roto_ptrs.txt contain mapping between the summary token and the corresponding matching token in the content plan. For eg: the last entry 245,39 in train_roto_ptrs[1] indicates that the 245th token in summary matches with 39th content plan entry.

Having said that, in later experiments I have realized that such supervision is not strictly required. The model learns an accurate value of p_copy even without such a supervision.
Hence I would recommend that you could comment out any code which uses the supervision through pointers and the model should work well enough.

[wingedRuslan]

Hi Ratish,

Thanks a lot for your prompt answer!

It was such a useful and valuable comment, that I was able to create same file but now for other dataset!

The entries in the train_roto_ptrs.txt contain mapping between the summary token and the corresponding matching token in the content plan. For eg: the last entry 245,39 in train_roto_ptrs[1] indicates that the 245th token in summary matches with 39th content plan entry.

At this moment I did not try to comment out any code which uses the supervision through pointers, since I want to keep things just the way you did! Later I plan to run such an experiment and see whether such supervision is required or not.

[ratishsp]

Great!
All the best.

[happycjksh]

When I drive the utils_data.py, the terminal output that the index out of range in the line 593. Do you have the problem?