PromptContrastiveDecoding/mt_task.py at main · xiaojing29/PromptContrastiveDecoding · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
import logging
import subprocess
import tempfile
import random
import copy
from pathlib import Path
from scripts.utils_run import FLORES101_CONVERT
from sacrebleu import get_source_file
from datasets import load_dataset
from tqdm import tqdm
import os
from translation_models.prompts import POSITIVE_PROMPTS, NEGATIVE_PROMPTS


class MTTask:

    def __init__(self,
                 src_lang: str,
                 tgt_lang: str,
                 testset: str,
                 ):
        self.src_lang = src_lang
        self.tgt_lang = tgt_lang
        self.language_pair = f"{src_lang}-{tgt_lang}"
        self.testset = testset
        base_out_dir = Path(__file__).parent / "out"
        print(base_out_dir)
        base_out_dir.mkdir(parents=True, exist_ok=True)
        assert base_out_dir.exists()
        self.out_dir = base_out_dir / self.testset
        self.out_dir.mkdir(exist_ok=True)

        self.out_dir = self.out_dir / self.language_pair
        self.out_dir.mkdir(exist_ok=True)
        self.load_converter = FLORES101_CONVERT

    def __str__(self):
        return f"{self.testset}-{self.src_lang}-{self.tgt_lang}"

    def evaluate(self, translation_method: callable, type='direct', source_contrastive=1, source_weight=None,language_contrastive=None, language_weight=None, prompt_contrastive=False, prompt_weight=None) -> Path:

        ## load FLORES dataset
        #source_sentences = load_dataset('gsarti/flores_101',self.load_converter[self.src_lang])['devtest']['sentence']

        ## load FLORES Plus dataset
        dataset = load_dataset('openlanguagedata/flores_plus')
        # Filter the 'dev' split for the target language using iso_639_3
        source_sentences = [
            example['text']
            for example in dataset['devtest']
            if example['iso_639_3'] == self.load_converter[self.src_lang]
        ]

        if type == 'direct':
            translations = translation_method(
            src_lang=self.src_lang,
            tgt_lang=self.tgt_lang,
            source_sentences=source_sentences,
            )
        elif type == 'contrastive':
            multi_source_sentences = []
            src_weights = [1]
            tgt_langs=[]
            src_langs=[]

            # randomly shuffled input to suppress hallucinations
            if source_contrastive:
                multi_source_sentences.append(source_sentences)
                src_weights.append(1.0)
                tgt_langs.append(self.tgt_lang)
                src_langs.append(self.src_lang)

                for i in range(source_contrastive):
                    shuffled_sentences = copy.copy(source_sentences)
                    random.shuffle(shuffled_sentences)
                    multi_source_sentences.append(shuffled_sentences)
                    src_weights.append(source_weight/source_contrastive)
                    tgt_langs.append(self.tgt_lang)
                    src_langs.append(self.src_lang)

                print(f"Source Weights (source contrastive): {src_weights}")

            # input with wrong target language indicator to suppress off-target translation
            if language_contrastive:
                multi_source_sentences.append(source_sentences)
                src_weights.append(1.0)
                tgt_langs.append(self.tgt_lang)
                src_langs.append(self.src_lang)

                for offtarget in language_contrastive:
                    # ignore contrastive variants that are identical to true translation direction
                    if offtarget == self.tgt_lang:
                        continue
                    # don't create contrastive variant for src language if language is already listed (avoid duplicates)
                    if offtarget == 'src' and self.src_lang in language_contrastive:
                        continue
                    multi_source_sentences.append(source_sentences)
                    src_weights.append(language_weight)
                    if offtarget == 'src':
                        tgt_langs.append(self.src_lang)
                    else:
                        tgt_langs.append(offtarget)
                    src_langs.append(self.src_lang)

                print(f"Language Weights (language contrastive): {src_weights}")


            # prompt-contrastive: add positive and negative prompts
            if prompt_contrastive:

                # Add positive prompts
                for positive_prompt in POSITIVE_PROMPTS:
                    #print(f"Adding positive prompt: {positive_prompt}")
                    positive_prompts = [
                        f"{src_sent}\n\n{positive_prompt}"
                        for src_sent in source_sentences
                    ]

                    multi_source_sentences.append(positive_prompts)
                    src_weights.append(1.0)  # Positive prompts have weight 1.0
                    tgt_langs.append(self.tgt_lang)
                    src_langs.append(self.src_lang)

                # Add negative prompts
                for negative_prompt in NEGATIVE_PROMPTS:
                    #print(f"Adding negative prompt: {negative_prompt}")
                    negative_prompts = [
                        f"{src_sent}\n\n{negative_prompt}"
                        for src_sent in source_sentences
                    ]

                    multi_source_sentences.append(negative_prompts)
                    src_weights.append(prompt_weight)  # Negative prompts have weight prompt_weight
                    tgt_langs.append(self.tgt_lang)
                    src_langs.append(self.src_lang)

                is_prompt_contrastive=True
                print(f"Prompt Weights: {src_weights}")

            else:
                is_prompt_contrastive=False

            translations = []
            for pair in tqdm(list(zip(*multi_source_sentences))):
                translation = translation_method(
                    src_langs=src_langs,
                    tgt_langs=tgt_langs,
                    src_weights=src_weights,
                    multi_source_sentences=pair,
                    is_prompt_contrastive=is_prompt_contrastive,
                    )
                translations.append(translation)
        else:
            raise NotImplementedError

        if type == 'direct':
            file_name = 'direct_devtest'
        elif type == 'contrastive':
            file_name = 'contrastive-{0}-{1}'.format(source_contrastive, source_weight)
            if language_contrastive:
                file_name += "-lang-{0}-{1}".format('+'.join(language_contrastive), language_weight)
            if prompt_contrastive:
                file_name = "contrastive-prompt-general".format(prompt_weight)
        else:
            raise NotImplementedError

        self.out_dir.mkdir(parents=True, exist_ok=True)

        with open(str(self.out_dir)+"/"+file_name+".txt", 'w', encoding='utf-8') as f:
            f.write("\n".join(translations))

        if not os.path.isfile(str(self.out_dir)+"/"+"ref.txt"):
            # target_sentences = load_dataset('gsarti/flores_101', self.load_converter[self.tgt_lang])['devtest']['sentence']
            target_sentences = [
                example['text']
                for example in dataset['devtest']
                if example['iso_639_3'] == self.load_converter[self.tgt_lang]
            ]

            with open(str(self.out_dir) + "/" + "ref.txt", 'w', encoding='utf-8') as f:
                f.write("\n".join(target_sentences))

        return Path(f.name)