shmsw25 / decomprc Goto Github PK

Thanks for the great work and for open-sourcing the code :) I'm seeing some (empty) directories in the "DecompRC-all-models-and-data.zip" file (under the "__MACOSX" directory). I'm guessing these files are unnecessary but I just wanted to confirm? Also to confirm that there aren't some files I'm missing. Thanks!

Hi,thanks for your open source code. In Decomposition Model, how does the predict_file data/decomposition-data/decomposition-bridge-dev-v1.json come? Although I know you have released the file , I want to know how do you preprocess the original data and obtain this file? Could you share the preprocess code?

I believe that project of this size, that has this many dependencies, should include a file (requirements.txt) in python, of the requirements to run this project.

Hi, thank you for your repository.

In my understanding, to obtain the scores for the comparison questions, we need to do the following steps:

1. Prepare sub-questions data

Download your code and data here, then run it, then prepare two files for the single-hop models (dev_c.1.json and dev_c.2.json).

2. Single-hop RC Model

Run the single-hop models for these files. For example:

python3 main.py --do_predict --output_dir out/compare_1 \
--predict_file data/decomposed/dev_c.1.json \
--init_checkpoint model/onehop/model1.pt,model/onehop/model2.pt,model/onehop/model3.pt \
--n_best_size 4

3. Obtain the final answer for each sample

From your code, I think we have 8 operators. I can use some heuristics to obtain the final results. If it is ok, could you share your file on obtaining the final score here for reproducing the results?

Also, please correct me if my understanding about your model is not correct.

Thank you for your support.

When I used convert_hotpot2squad.py to convert HotpotQA into SQuAD style,the error 'ImportError: cannot import name 'find_span' 'would be occurred. I did not find 'find_span' in the prepro_util.py, there is only find_span_from_text

What is the difference between y_no and y_none mentioned in Section 3.3 of the paper when calculating scores?

Thank you very much for sharing your results. Can you tell me how to get the file DecompRC-all-models-and-data.zip?

I see that in the code and README you use "hotpot_train_v1.json". However, when I download the training set from the HotpotQA website, I receive "hotpot_train_v1.1.json". Is there a difference between these two versions? If so, what are the differences, and which version did you use?

I was able to run the code for Distractor settings however when I try to do the same for fullwiki the code expects the file dev_distractor only ?

How can I run the same for Dev_fullwiki ?

Could you please explain how the type_dev_predictions.json file in the open-source code of your paper 'Multi-hop Reading Comprehension through Question Decomposition and Rescoring' was generated? Thank you.

this comand seems needed to be update

in the #intersection，there is no file is "/hotpot_decomposed/annotated-span-intersec-dev-v1.1.json "

# Intersection
python3 main.py --do_train --model span-predictor --output_dir out/decom-intersec \
              --predict_file /home/sewon/data/hotpot_decomposed/annotated-span-intersec-dev-v1.1.json \
              --train_file /home/sewon/data/hotpot_decomposed/annotated-span-intersec-train-v1.1.json \
              --max_seq_length 100 --train_batch_size 50 --predict_batch_size 100 \
              --max_n_answers 1 --eval_period 50 --num_train_epochs 5000 --wait_step 50

do you mean #intersection should write like this? the same directory as #Bridge?

#Intersection
python3 main.py --do_train --model span-predictor --output_dir out/decom-bridge \
              --train_file data/decomposition-data/decomposition-intersec-train-v1.json \
              --predict_file data/decomposition-data/decomposition-intersec-dev-v1.json \
              --max_seq_length 100 --train_batch_size 50 --predict_batch_size 100 \
              --max_n_answers 1 --eval_period 50 --num_train_epochs 5000 --wait_step 50 --with_key

when I run the command python main.py --do_train --output_dir out/hotpot --train_file data/hotpot-all/train.json --predict_file data/hotpot-all/dev.json errors occurs : joblib.externals.loky.process_executor.BrokenProcessPool: A task has failed to un-serialize. Please ensure that the arguments of the function are all picklable.
Maybe it is due to different version of joblib ? Could you tell me the version of your joblib in the task?
my joblib version is 0.14.1

Hi,

Thank you for your work and repository.

I'm trying to use the type classifier explained in your paper in Sec. 3.4 (Pipeline Approach), where you first predict the reasoning type of the question. You use a BERT model with a 4-way classifier on top. However, examining the type classifier model file released in "DecompRC-all-models-and-data.zip", the classifier is 2-way (similar to the scorer model). Am I missing something? Is the "type-classifier" folder released different than the reasoning type classifier explained in the paper?

Since only the type predictions of dev are released, can you please either release the train questions type predictions or the classifier model?

Thanks in advance!

Hey,

Do you know when the comparison code will be released? I am really interested the model finds the correct span and the evidence that is uses for support.

Thank you for your work on this!

Cheers,
Mitch