Logic2Text

Data

In the dataset folder, we have the full dataset (all_data.json), and the train test split (train.json, valid.json, test.json). Each example is in a dictionary of the following format:

    "topic": table caption,
    "wiki": link to the wikipedia page,
    "action": logic type,
    "sent": description,
    "annotation": raw annotation data from the mturk,
    "logic": logic form,
    "logic_str": linearized logic form,
    "nid": number of fuction nodes,
    "table_header": table header, in list format
    "table_cont": table content, list of lists
  

Logic form execution

In the execution folder, run

python execute.py

It will execute all the logic forms in all_data.json. All the function definitions are in APIs.py

This site is under construction, and we will release other codes in the future.

Model

The pre-trained GPT-2 can be downloaded via Dropbox.

requirements

python 3.6 tensorflow 1.12

data pre-process

In addition to the original data folder, this script will create another folder to contain the preprocessed data, which will be used for train and test.

python preprocess.py data_folder GPT_folder

train

Modify the data paths and parameters in the Main.py file. Then run:

CUDA_VISIBLE_DEVICES=0,1 python Main.py --mode train

test

CUDA_VISIBLE_DEVICES=0,1 python Main.py --mode test --saved_model_path load_model_path

AMT annotation interface

In the folder amt, we have all the html scripts for AMT annotations.

folder composition: html files for statements composition
folder logic_form: html files for logic form annotation
folder quality_check: html files for checking logic form annotation quality