File size: 14,974 Bytes
3358886
 
 
 
8b87b6e
fc63a12
 
 
3358886
fc63a12
8b87b6e
81c18c4
8b87b6e
1db0637
3358886
e735225
 
3358886
fc63a12
 
81c18c4
 
3358886
fc63a12
 
3358886
fc63a12
8b87b6e
 
3358886
e735225
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8b87b6e
fc63a12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3358886
 
090aa41
 
 
 
fc63a12
 
 
090aa41
 
fc63a12
 
 
 
 
 
 
 
 
 
3358886
 
 
 
 
 
 
 
 
fc63a12
3358886
fc63a12
81c18c4
fc63a12
 
3358886
8b87b6e
3358886
 
 
 
 
81c18c4
fc63a12
3358886
fc63a12
 
3358886
fc63a12
3358886
 
e735225
 
3358886
e735225
3358886
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
81c18c4
e735225
81c18c4
e735225
 
 
 
 
 
81c18c4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
965aebe
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3358886
 
81c18c4
 
 
 
 
 
 
 
 
 
 
 
 
3358886
965aebe
1db0637
81c18c4
 
 
 
2ff0c49
1db0637
 
3358886
1db0637
 
 
 
3358886
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fc63a12
3358886
 
 
 
 
 
 
 
 
 
fc63a12
3358886
 
 
 
 
 
e735225
3358886
e735225
 
3358886
fc63a12
3358886
 
e735225
 
 
 
 
 
3358886
 
8b87b6e
3358886
 
 
 
 
 
 
 
e735225
3358886
 
81c18c4
fc63a12
 
3358886
 
e735225
3358886
 
 
 
 
965aebe
 
 
 
 
 
 
 
3358886
 
 
fc63a12
3358886
 
 
 
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
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
import os
import re
import random
import requests
import gradio as gr
import numpy as np
import nltk 
from newspaper import Article
from nltk.tokenize import sent_tokenize
from sentence_transformers import SentenceTransformer, util
import spacy
import nltkmodule
import en_core_sci_lg
from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
import torch

#nltk.download('punkt')

# --- Models (load once, globally) ---
scispacy = en_core_sci_lg.load()
sbert_keybert = SentenceTransformer("pritamdeka/BioBERT-mnli-snli-scinli-scitail-mednli-stsb")
sbert_rerank = SentenceTransformer("pritamdeka/S-PubMedBert-MS-MARCO")
NLI_MODEL_NAME = "pritamdeka/PubMedBERT-MNLI-MedNLI"
nli_tokenizer = AutoTokenizer.from_pretrained(NLI_MODEL_NAME)
nli_model = AutoModelForSequenceClassification.from_pretrained(NLI_MODEL_NAME)
NLI_LABELS = ['CONTRADICTION', 'NEUTRAL', 'ENTAILMENT']

PUBMED_N = 100
TOP_ABSTRACTS = 10

# --- Sentence section classifier model (BioBert-PubMed200kRCT) ---
EVIDENCE_MODEL = "pritamdeka/BioBert-PubMed200kRCT"
evidence_tokenizer = AutoTokenizer.from_pretrained(EVIDENCE_MODEL)
evidence_model = AutoModelForSequenceClassification.from_pretrained(EVIDENCE_MODEL)
label_map = {0: "BACKGROUND", 1: "OBJECTIVE", 2: "METHODS", 3: "RESULTS", 4: "CONCLUSIONS"}

def extract_evidence_sentences_from_abstract(abstract, keep_labels=("RESULTS", "CONCLUSIONS")):
    sents = sent_tokenize(abstract)
    evidence_sents = []
    for s in sents:
        inputs = evidence_tokenizer(s, return_tensors="pt", truncation=True, padding=True)
        with torch.no_grad():
            logits = evidence_model(**inputs).logits
            pred = torch.argmax(logits, dim=1).item()
            label = label_map[pred]
            if label in keep_labels:
                evidence_sents.append((label, s))
    return evidence_sents

# --- Europe PMC retrieval ---
def retrieve_europepmc_abstracts_simple(text, n=TOP_ABSTRACTS):
    query = get_keybert_query(text, top_n=7)
    print("Trying Europe PMC query:", query)
    url = f'https://www.ebi.ac.uk/europepmc/webservices/rest/search?query={query}&resulttype=core&format=json&pageSize={n}'
    r = requests.get(url)
    results = r.json().get('resultList', {}).get('result', [])
    titles = [res.get('title', '') for res in results]
    abstracts = [res.get('abstractText', '') for res in results]
    return titles, abstracts

# --- Utility: get robust keybert-style query ---
def get_keybert_query(text, top_n=10):
    doc = scispacy(text)
    phrases = [ent.text for ent in doc.ents]
    if not phrases:
        phrases = [chunk.text for chunk in doc.noun_chunks]
    phrases = list(set([ph.strip() for ph in phrases if len(ph) > 2]))
    if not phrases:
        return ""
    doc_emb = sbert_keybert.encode([text])
    phrase_embs = sbert_keybert.encode(phrases)
    sims = np.array(util.pytorch_cos_sim(doc_emb, phrase_embs))[0]
    top_idxs = sims.argsort()[-top_n:]
    keywords = [phrases[i] for i in top_idxs]
    query = " OR ".join(f'"{kw}"' for kw in keywords)
    return query

# --- Claim extraction ---
indicator_phrases = [
    "found that", "findings suggest", "shows that", "showed that", "demonstrated", "demonstrates",
    "revealed", "reveals", "suggests", "suggested", "indicated", "indicates", "reported", "reports",
    "was reported", "concluded", "concludes", "conclusion", "authors state", "stated", "data suggest",
    "observed", "observes", "study suggests", "study shows", "study found", "researchers found",
    "results indicate", "results show", "confirmed", "confirm", "confirming", "point to",
    "documented", "document", "evidence of", "evidence suggests", "associated with", "correlated with",
    "link between", "linked to", "relationship between", "was linked", "connected to", "relationship with",
    "tied to", "association with", "increase", "increases", "increased", "decrease", "decreases", "decreased",
    "greater risk", "lower risk", "higher risk", "reduced risk", "raises the risk", "reduces the risk",
    "risk of", "risk for", "likelihood of", "probability of", "chance of", "rate of", "incidence of",
    "prevalence of", "mortality", "survival rate", "death rate", "odds of", "number of", "percentage of",
    "percent of", "caused by", "causes", "cause", "resulted in", "results in", "leads to", "led to",
    "contributed to", "responsible for", "due to", "as a result", "because of",
    "randomized controlled trial", "RCT", "clinical trial", "participants", "enrolled", "sample size",
    "statistically significant", "compared to", "compared with", "versus", "compared against",
    "more than", "less than", "greater than", "lower than", "higher than", "significantly higher",
    "significantly lower", "significantly increased", "significantly decreased", "significant difference",
    "effect of", "impact of", "influence of", "predictor of", "predicts", "predictive of", "factor for",
    "determinant of", "plays a role in", "contributes to", "related to", "affects", "influences",
    "difference between", "according to", "a recent study", "researchers from"
]

def extract_claims_pattern(article_text):
    sentences = sent_tokenize(article_text)
    claims = [
        s for s in sentences
        if any(phrase in s.lower() for phrase in indicator_phrases)
        or re.search(r"\b\d+(\.\d+)?%?\b", s)
    ]
    return list(dict.fromkeys(claims))

def match_claims_to_headline(claims, headline):
    emb_model = sbert_keybert
    headline_emb = emb_model.encode([headline])
    claim_embs = emb_model.encode(claims)
    sims = util.pytorch_cos_sim(headline_emb, claim_embs)[0]
    matched_claims = [claim for claim, sim in zip(claims, sims) if sim >= 0.6]
    if not matched_claims and claims:
        idxs = np.argsort(-sims.cpu().numpy())[:min(3, len(claims))]
        matched_claims = [claims[i] for i in idxs]
    return matched_claims

# --- Semantic reranking ---
def semantic_rerank_claim_abstracts(claim, titles, abstracts, top_k=TOP_ABSTRACTS):
    doc_texts = [f"{t}. {a}" for t, a in zip(titles, abstracts)]
    doc_embs = sbert_rerank.encode(doc_texts)
    claim_emb = sbert_rerank.encode([claim])
    sims = util.pytorch_cos_sim(claim_emb, doc_embs)[0]
    idxs = np.argsort(-sims.cpu().numpy())[:top_k]
    return [titles[i] for i in idxs], [abstracts[i] for i in idxs]

# --- NLI evidence extraction (run only on results/conclusion sentences) ---
def extract_evidence_nli(claim, evidence_sentences):
    evidence = []
    for sent in evidence_sentences:
        encoding = nli_tokenizer(
            sent, claim,
            return_tensors='pt',
            truncation=True,
            max_length=256,
            padding=True
        )
        with torch.no_grad():
            outputs = nli_model(**encoding)
            probs = torch.softmax(outputs.logits, dim=1).cpu().numpy().flatten()
            max_idx = probs.argmax()
            label = NLI_LABELS[max_idx]
            score = float(probs[max_idx])
        evidence.append({
            "sentence": sent,
            "label": label,
            "score": score
        })
    return evidence

# --- Summarizer model options (now with Mistral API!) ---
model_options = {
    "Mistral Small (API, fast/free)": "mistral-small-2503",
    "Llama-3.2-1B-Instruct (Meta, gated)": "meta-llama/Llama-3.2-1B-Instruct",
    "Gemma-3-1B-it (Google, gated)": "google/gemma-3-1b-it",
    "TinyLlama-1.1B-Chat (Open)": "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
}
pipe_cache = {}

# --- Mistral API summarization ---
def summarize_with_mistral_api(prompt, model_name="mistral-small", max_tokens=128, temperature=0.1):
    api_key = os.getenv("MISTRAL_API_KEY")
    if not api_key:
        return "Missing MISTRAL_API_KEY secret/env variable!"
    endpoint = "https://api.mistral.ai/v1/chat/completions"
    headers = {
        "Authorization": f"Bearer {api_key}",
        "Content-Type": "application/json"
    }
    data = {
        "model": model_name,
        "messages": [{"role": "user", "content": prompt}],
        "max_tokens": max_tokens,
        "temperature": temperature
    }
    response = requests.post(endpoint, headers=headers, json=data, timeout=30)
    if response.status_code == 200:
        content = response.json()["choices"][0]["message"]["content"]
        return content.strip()
    else:
        return f"API Error ({response.status_code}): {response.text}"

def get_summarizer(model_choice):
    model_id = model_options[model_choice]
    if model_id in pipe_cache:
        return pipe_cache[model_id]
    kwargs = {
        "model": model_id,
        "torch_dtype": torch.float16 if torch.cuda.is_available() else torch.float32,
        "device_map": "auto",
        "max_new_tokens": 128
    }
    if any(gated in model_id for gated in ["meta-llama", "gemma"]):
        hf_token = os.environ.get("HF_TOKEN", None)
        if hf_token:
            kwargs["token"] = hf_token
        else:
            raise RuntimeError(f"Model '{model_choice}' requires a Hugging Face access token. Please set 'HF_TOKEN' as a Space secret or environment variable.")
    pipe_cache[model_id] = pipeline("text-generation", **kwargs)
    return pipe_cache[model_id]

def summarize_evidence_llm(claim, evidence_list, model_choice):
    support = [ev['sentence'] for ev in evidence_list if ev['label'] == 'ENTAILMENT']
    contradict = [ev['sentence'] for ev in evidence_list if ev['label'] == 'CONTRADICTION']
    user_prompt = (
        f"Claim: {claim}\n"
        f"Supporting evidence:\n" + ("\n".join(support) if support else "None") + "\n"
        f"Contradicting evidence:\n" + ("\n".join(contradict) if contradict else "None") + "\n"
        "Explain to a layperson: Is this claim likely true, false, or uncertain based on the evidence above? Give a brief and simple explanation in 2-3 sentences."
    )
    if model_choice in [
        "Mistral Small (API, fast/free)",
        "Mistral Medium (API, free tier)",
        "Mistral Large (API, may require paid)"
    ]:
        mistral_name = model_options[model_choice]
        return summarize_with_mistral_api(user_prompt, model_name=mistral_name)
    try:
        pipe = get_summarizer(model_choice)
        outputs = pipe(
            [
                {"role": "system", "content": "You are a helpful biomedical assistant. Summarize scientific evidence in plain English for the general public."},
                {"role": "user", "content": user_prompt}
            ],
            max_new_tokens=128,
            do_sample=False,
            temperature=0.1,
        )
        out = outputs[0]["generated_text"]
        if isinstance(out, list) and "content" in out[-1]:
            return out[-1]["content"].strip()
        return out.strip()
    except Exception as e:
        return f"Summary could not be generated: {e}"

def format_evidence_html(evidence_list):
    color_map = {"ENTAILMENT":"#e6ffe6", "CONTRADICTION":"#ffe6e6", "NEUTRAL":"#f8f8f8"}
    html = ""
    for ev in evidence_list:
        color = color_map[ev["label"]]
        html += (
            f'<div style="background:{color};padding:6px;border-radius:6px;margin-bottom:3px">'
            f'<b>{ev["label"]}</b> (confidence {ev["score"]:.2f}): {ev["sentence"]}'
            '</div>'
        )
    return html

def factcheck_app(article_url, summarizer_choice):
    try:
        art = Article(article_url)
        art.download()
        art.parse()
        text = art.text
        headline = art.title
    except Exception as e:
        return f"<b>Error downloading or reading article:</b> {e}", None

    claims = extract_claims_pattern(text)
    matched_claims = match_claims_to_headline(claims, headline)
    if not matched_claims:
        return "<b>No check-worthy claims found that match the headline.</b>", None

    results_html = ""
    all_results = []
    for claim in matched_claims:
        titles, abstracts = retrieve_europepmc_abstracts_simple(claim)
        if not titles:
            results_html += f"<hr><b>Claim:</b> {claim}<br><i>No relevant abstracts found in Europe PMC.</i><br>"
            all_results.append({"claim": claim, "summary": "No abstracts found.", "evidence": []})
            continue
        top_titles, top_abstracts = semantic_rerank_claim_abstracts(claim, titles, abstracts)
        evidence_results = []
        for title, abstract in zip(top_titles, top_abstracts):
            ev_sents = extract_evidence_sentences_from_abstract(abstract)
            if ev_sents:
                sent_list = [s for lbl, s in ev_sents]
            else:
                sent_list = sent_tokenize(abstract)
            evidence = extract_evidence_nli(claim, sent_list)
            evidence_results.append({"title": title, "evidence": evidence})
        all_evidence_sentences = [ev for abs_res in evidence_results for ev in abs_res["evidence"]]
        summary = summarize_evidence_llm(claim, all_evidence_sentences, summarizer_choice)
        results_html += f"<hr><b>Claim:</b> {claim}<br><b>Layman summary:</b> {summary}<br>"
        for abs_res in evidence_results:
            results_html += f"<br><b>Abstract:</b> {abs_res['title']}<br>{format_evidence_html(abs_res['evidence'])}"
        all_results.append({"claim": claim, "summary": summary, "evidence": evidence_results})
    return results_html, all_results

description = """
<b>What does this app do?</b><br>
This app extracts key scientific claims from a news article, finds the most relevant biomedical research papers using robust keyphrase extraction and semantic reranking, checks which sentences in those papers support or contradict each claim, and gives you a plain-English summary verdict.<br><br>
<b>How to use it:</b><br>
1. Paste the link to a biomedical news article.<br>
2. Choose an AI summarizer model below. If you have no special access, use 'TinyLlama' or 'Mistral Small' (works for everyone, free).<br>
3. Wait for the results.<br>
4. For each claim, you will see:<br>
- A plain summary of what research says.<br>
- Color-coded evidence sentences (green=support, red=contradict, gray=neutral).<br>
- The titles of the most relevant research articles.<br><br>
<b>Everything is 100% open source and runs on this website—no personal info or cloud API needed.</b>
"""

iface = gr.Interface(
    fn=factcheck_app,
    inputs=[
        gr.Textbox(lines=2, label="Paste a news article URL"),
        gr.Dropdown(
            choices=list(model_options.keys()),
            value="TinyLlama-1.1B-Chat (Open)",
            label="Choose summarizer model"
        )
    ],
    outputs=[gr.HTML(label="Fact-Check Results (Summary & Evidence)"), gr.JSON(label="All Results (JSON)")],
    title="BioMedical News Fact-Checking & Research Evidence Finder",
    description=description,
    examples=[["https://www.medicalnewstoday.com/articles/omicron-what-do-we-know-about-the-stealth-variant", "TinyLlama-1.1B-Chat (Open)"]],
    allow_flagging="never"
)

iface.launch(share=False, server_name='0.0.0.0', show_error=True)