create_multilingual_simCSE.py

import json
import jsonlines
import argparse
import pandas as pd
import pprint
from random import *

def main(args):
    mainDict = {}
    with jsonlines.open(args.norwegian_input_file) as reader:
        for obj in reader:
            #Check if the value already exists
            neutral = ""
            entailment = ""
            contradiction = ""
            prompt = ""

            if mainDict.get(obj['promptID'], None):
                prompt = obj['sentence1']
                entailment = mainDict[obj['promptID']].get('entailment','')
                contradiction = mainDict[obj['promptID']].get('contradiction','')
                neutral = mainDict[obj['promptID']].get('neutral','')

            if obj['gold_label'] == "neutral":
                neutral = obj['sentence2']
            elif obj['gold_label'] == "contradiction":
                contradiction =  obj['sentence2']
            elif obj['gold_label'] == "entailment":
                entailment = obj['sentence2']
            
            mainDict[obj['promptID']] = {'prompt': prompt, 'entailment' : entailment, 'neutral' : neutral, 'contradiction' : contradiction}
    
    myList = []
    for promptID in mainDict:
        myList.append({'sent0':mainDict[promptID]['prompt'], 'sent1':mainDict[promptID]['entailment'], 'hard_neg':mainDict[promptID]['contradiction']})
    
    df = pd.DataFrame.from_records(myList)        
    
    #Drop empty
    df.replace("", float("NaN"), inplace=True)
    df.dropna(subset = ["sent0","sent1","hard_neg"], inplace=True)
    
    #Get the english file
    english = pd.read_csv(args.english_input_file)

    #Get the multilingual lookup file
    multi = pd.read_csv(args.multilingual_input_file, sep='\t')
    multi.columns=['id','english','norwegian']
    multi['english'] = multi['english'].str.strip()
    multi['norwegian'] = multi['norwegian'].str.strip()
    
    #Create dictionary
    mydict = {}
    for index,row in multi.iterrows():
        mydict[row['norwegian']] = row['english']

    merged = df.copy()
    for index, row in merged.iterrows():
        flip = randint(1,6)
        if flip == 3 or flip == 4 or flip == 5:
            merged.at[index, 'sent0'] = mydict.get(row['sent0'],row['sent0'])
        
        if flip == 2 or flip == 4 or flip == 6:
            merged.at[index, 'sent1'] = mydict.get(row['sent1'],row['sent1'])
        
        if flip == 1 or flip == 5 or flip == 6:
            merged.at[index,'hard_neg'] = mydict.get(row['hard_neg'],row['hard_neg'])
        
    
    combined = pd.concat([df, english, merged])
    
    combined = combined.sample(frac=1).reset_index(drop=True)

    #Save csv
    combined.to_csv(args.output_file, index=False)

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--norwegian_input_file', help="Norwegian json file.", required=True)
    parser.add_argument('--english_input_file', help="English csv file.", required=True)
    parser.add_argument('--multilingual_input_file', help="Multilingual tsv file.", required=True)
    parser.add_argument('--output_file', help="Output file.", required=True)
    
    args = parser.parse_args()
    main(args)