File size: 9,924 Bytes
5c3b4a6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
from transformers import pipeline, AutoTokenizer
import unittest
from mediaunmasked.scrapers.article_scraper import ArticleScraper
from tabulate import tabulate
import torch
from typing import List
import logging
import transformers

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

class MediaUnmaskLLMTester(unittest.TestCase):
    transformers.logging.set_verbosity_error()
    def setUp(self):
        """Set up LLMs and scrape article."""
        self.models = {
            # Upgraded Evidence-Based Models
            "RoBERTa-MNLI": {"model": "roberta-large-mnli", "max_length": 512},  # Corrected to standard MNLI model
            "DeBERTa-Fact": {"model": "MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli", "max_length": 512},
            "T5-Large": {"model": "google/t5-v1_1-large", "max_length": 512},
            "SciBERT": {"model": "allenai/scibert_scivocab_uncased", "max_length": 512},
            "BART-FEVER": {"model": "facebook/bart-large", "max_length": 1024},  # Note: Needs FEVER fine-tuning
            "MultiQA-MiniLM": {"model": "sentence-transformers/multi-qa-MiniLM-L6-cos-v1", "max_length": 512},

            # Existing Models for Benchmarking
            "BART-MNLI": {"model": "facebook/bart-large-mnli", "max_length": 1024},
            "RoBERTa-Bias": {"model": "cardiffnlp/twitter-roberta-base-hate", "max_length": 512},
            "DistilBERT-Sentiment": {"model": "distilbert-base-uncased-finetuned-sst-2-english", "max_length": 512},
            "GPT2-Generation": {"model": "gpt2", "max_length": 1024},
        }

        self.device = 0 if torch.cuda.is_available() else -1
        self.scraper = ArticleScraper()
        self.article_url = "https://www.snopes.com/fact-check/trump-super-bowl-cost-taxpayers/"
        self.article_data = self.scraper.scrape_article(self.article_url) or {}

        self.results = {
            "headline": self.article_data.get("headline", "No headline"),
            "content": self.article_data.get("content", "No content available"),
            "scores": {}
        }

        self.tokenizers = {name: AutoTokenizer.from_pretrained(model["model"]) for name, model in self.models.items()}

    def _split_content(self, model_name: str, content: str) -> List[str]:
        """Split content into sections within model token limits, ensuring valid output."""
        tokenizer = self.tokenizers[model_name]
        max_length = self.models[model_name]["max_length"]

        if not content or not content.strip():
            return ["No valid content"]

        encoded = tokenizer.encode_plus(content, add_special_tokens=True, truncation=True, max_length=max_length)
        decoded = tokenizer.decode(encoded["input_ids"], skip_special_tokens=True)
        
        return [decoded] if decoded.strip() else ["No valid content"]

    def _get_flagged_phrases(self, model_pipeline, sections, threshold=0.6, top_k=5):
        """Extract top-scoring flagged phrases while handling None values safely."""
        if not sections or not isinstance(sections, list):
            return [("None", "N/A")]

        flagged_phrases = []
        
        for section in sections:
            if not section or not isinstance(section, str) or not section.strip():  # Ensure section is a valid string
                continue

            sentences = [s.strip() for s in section.split(". ") if s.strip()]
            for sentence in sentences:
                if not sentence or not isinstance(sentence, str):  # Double-check before running the model
                    continue

                try:
                    preds = model_pipeline(sentence)
                    if preds and isinstance(preds, list):
                        top_pred = max(preds, key=lambda x: x["score"])
                        if top_pred["score"] >= threshold:
                            short_phrase = " ".join(sentence.split()[:10])  # Shorten for readability
                            flagged_phrases.append((short_phrase, top_pred["score"], top_pred["label"]))
                except Exception as e:
                    logger.error(f"Error analyzing sentence: {e}")
                    continue

        flagged_phrases.sort(key=lambda x: x[1], reverse=True)
        return [(phrase, label) for phrase, _, label in flagged_phrases[:top_k]] or [("None", "N/A")]

    def test_headline_vs_content(self):
        """Check headline-content alignment."""
        headline = self.results["headline"]
        content = self.results["content"]

        for model_name in self.models:
            with self.subTest(model=model_name):
                analyzer = pipeline("text-classification", model=self.models[model_name]["model"], device=self.device)
                sections = self._split_content(model_name, content)

                headline_score = max(analyzer(headline), key=lambda x: x["score"])["score"]
                content_scores = [max(analyzer(section), key=lambda x: x["score"])["score"] for section in sections]
                avg_content_score = sum(content_scores) / len(content_scores)
                consistency_score = abs(headline_score - avg_content_score)

                flagged_phrases = self._get_flagged_phrases(analyzer, sections)
                self.results["scores"].setdefault("headline_vs_content", {})[model_name] = {
                    "score": consistency_score,
                    "flagged_phrases": flagged_phrases
                }
                self.assertIsNotNone(consistency_score)

    def test_evidence_based(self):
        """Test evidence-based content."""
        content = self.results["content"]

        for model_name in self.models:
            if any(keyword in model_name.lower() for keyword in ["mnli", "fact", "fever", "qa"]):
                with self.subTest(model=model_name):
                    classifier = pipeline("zero-shot-classification", model=self.models[model_name]["model"], device=self.device)
                    sections = self._split_content(model_name, content)

                    results = [classifier(section, candidate_labels=["evidence-based", "opinion", "misleading"]) for section in sections]
                    avg_score = sum(r["scores"][r["labels"].index("evidence-based")] for r in results) / len(results)

                    flagged_phrases = self._get_flagged_phrases(classifier, sections)
                    self.results["scores"].setdefault("evidence_based", {})[model_name] = {
                        "score": avg_score,
                        "flagged_phrases": flagged_phrases
                    }
                    self.assertIsNotNone(avg_score)

    def test_manipulative_language(self):
        """Detect manipulative language."""
        content = self.results["content"]

        for model_name in self.models:
            if "sentiment" in model_name.lower() or "emotion" in model_name.lower() or "gpt" in model_name.lower():
                with self.subTest(model=model_name):
                    detector = pipeline("text-classification", model=self.models[model_name]["model"], device=self.device)
                    sections = self._split_content(model_name, content)

                    results = [max(detector(section), key=lambda x: x["score"]) for section in sections]
                    avg_score = sum(r["score"] for r in results) / len(results)

                    flagged_phrases = self._get_flagged_phrases(detector, sections)
                    self.results["scores"].setdefault("manipulative_language", {})[model_name] = {
                        "score": avg_score,
                        "flagged_phrases": flagged_phrases
                    }
                    self.assertIsNotNone(avg_score)

    def test_bias_detection(self):
        """Detect bias."""
        content = self.results["content"]

        for model_name in self.models:
            if "bias" in model_name.lower() or "toxic" in model_name.lower() or "roberta" in model_name.lower():
                with self.subTest(model=model_name):
                    detector = pipeline("text-classification", model=self.models[model_name]["model"], device=self.device)
                    sections = self._split_content(model_name, content)

                    results = [max(detector(section), key=lambda x: x["score"]) for section in sections]
                    avg_score = sum(r["score"] for r in results) / len(results)

                    flagged_phrases = self._get_flagged_phrases(detector, sections)
                    self.results["scores"].setdefault("bias_detection", {})[model_name] = {
                        "score": avg_score,
                        "flagged_phrases": flagged_phrases
                    }
                    self.assertIsNotNone(avg_score)

    def tearDown(self):
        """Print top 2 models per test with clearer formatting."""
        print("\n=== Top Model Recommendations ===")
        
        for test_type, model_results in self.results["scores"].items():
            print(f"\nTop 2 Models for {test_type}:")
            
            sorted_results = sorted(
                model_results.items(),
                key=lambda x: x[1]["score"],
                reverse=(test_type != "headline_vs_content")
            )
            
            top_2 = sorted_results[:2]
            table = [
                [
                    model,
                    f"{res['score']:.6f}",
                    ", ".join(f"{phrase} ({label})" for phrase, label in res["flagged_phrases"])
                ]
                for model, res in top_2
            ]

            print(tabulate(table, headers=["Model", "Score", "Flagged Phrases"], tablefmt="grid"))
            criteria = "Lowest consistency score (better alignment)" if test_type == "headline_vs_content" else "Highest detection score"
            print(f"Criteria: {criteria}")

if __name__ == "__main__":
    unittest.main()