Add backtranslation code

Code/backtranslate.py

+""" Backtranslation: Translate original sentences using Fairseq (https://github.com/facebookresearch/fairseq/blob/main/examples/translation/README.md)
+to create 5 paraphrases """
+
+import numpy as np
+import pandas as pd
+import json
+import torch
+import tqdm as notebook_tqdm
+import os 
+
+# Define backtranslation temperature parameter
+#temperature = 0.8
+temperature = 1.2
+
+def load_data_set(file_name):
+    with open(file_name, "r") as file:
+        data = file.read()
+        return json.loads(data)
+
+def load_data_sets(data_dir):
+    semeval_train_data = load_data_set(os.path.join(data_dir, "semeval_train.txt"))
+    companies_train_data = load_data_set(os.path.join(data_dir, "companies_train.txt"))
+    relocar_train_data = load_data_set(os.path.join(data_dir, "relocar_train.txt"))
+    return semeval_train_data, companies_train_data, relocar_train_data
+
+data_dir = "./data" 
+semeval_train_data, companies_train_data, relocar_train_data = load_data_sets(data_dir)
+
+
+#Load Fairseq transformers trained on WMT'19
+en2de = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.en-de.single_model', tokenizer='moses', bpe='fastbpe')
+de2en = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.de-en.single_model', tokenizer='moses', bpe='fastbpe')
+
+en2de.eval()  # disable dropout
+
+en2de.cuda() # move model to GPU for faster translation
+de2en.cuda()
+
+
+# Define helper functions
+def extract_target(dp):
+    start = dp["pos"][0]
+    end = dp["pos"][1]
+    target = dp["sentence"][start:end]
+    target = " ".join(target).lower()
+    return target
+
+def join_tokens(dp):
+    sent = " ".join(dp)
+    sent = sent.replace(' ,', ',')
+    sent = sent.replace(' .', '.')
+    sent = sent.replace(' !', '!')
+    sent = sent.replace(' ?', '?')
+    sent = sent.replace(' :', ':')
+    sent = sent.replace(' ;', ';')
+    sent = sent.replace(" '", "'")
+    sent = sent.replace(' "', '"')
+    sent = sent.replace('` ', '')
+    return sent
+
+# Apply backtranslation
+datasets = [semeval_train_data, companies_train_data, relocar_train_data]
+path_names = [f"semeval_loc_with_paraphrases_temp{temperature}", f"semeval_org_with_paraphrases_temp{temperature}", f"relocar_with_paraphrases_temp{temperature}"]
+
+for dataset, name in zip(datasets, path_names):
+    # Preprocess original sentences
+    data = pd.DataFrame.from_dict(dataset, orient='columns')
+
+    # Extract targets & add as new column
+    target_words = [extract_target(row) for index, row in data.iterrows()]
+    data["targets"] = target_words
+
+    # Join split original sentence to single sentence
+    joined_sents = [join_tokens(sent) for sent in data["sentence"]]
+    data["joined_sents"] = joined_sents
+
+    # Paraphrase joined original sentences
+    def paraphrase(sent, temperature):
+        en_encode = en2de.encode(sent)
+        outputs = en2de.generate(en_encode, sampling=True, topp=0.5, temperature=temperature)
+        nucleus = [en2de.decode(x['tokens']) for x in outputs]
+
+        multi_paraphrases = []      #backtranslate 4 times
+        for sentence in nucleus:
+            de_1 = de2en.translate(sentence)
+            de_2 = en2de.translate(de_1)
+            paraphrase = de2en.translate(de_2)
+            multi_paraphrases.append(paraphrase)
+
+        return multi_paraphrases
+
+    paraphrases = [paraphrase(sent, temperature) for sent in data["joined_sents"]]
+    data["paraphrases"] = paraphrases
+
+    data_path = os.path.join(data_dir, name)
+    data.to_csv(data_path)
+    print(f"Done: paraphrases generated and saved for {name}.")