import argparse import json import os import time from pathlib import Path from typing import List import tqdm from openai import OpenAI # ----------------------------- # SUBCLAIM EXTRACTION PROMPT # ----------------------------- def extraction_prompt(medical_text: str) -> str: prompt = f""" You are an expert medical annotator. Your task is to extract granular, factual subclaims from medical text. A subclaim is the smallest standalone factual unit that can be independently verified. Instructions: 1. Read the provided medical text. 2. Break it into clear, objective, atomic subclaims. 3. Each subclaim must come directly from the text. Do not infer or add information. 4. Keep subclaims short, non-overlapping, and de-duplicated. 5. Preserve numbers, units, and dates exactly as written. 6. If the text is empty, return an empty JSON list. 7. Return ONLY a valid JSON list of strings (no extra text). Medical Text: {medical_text} Return your output in JSON list format: [ "subclaim 1", "subclaim 2" ] """ return prompt def _load_openai_client() -> OpenAI: api_file = "/home/mshahidul/api_new.json" with open(api_file, "r") as f: api_keys = json.load(f) return OpenAI(api_key=api_keys["openai"]) def _parse_json_list(text: str) -> List[str]: cleaned = text.replace("```json", "").replace("```", "").strip() start_idx = cleaned.find("[") end_idx = cleaned.rfind("]") + 1 if start_idx == -1 or end_idx <= start_idx: raise ValueError("No JSON list found") parsed = json.loads(cleaned[start_idx:end_idx]) if not isinstance(parsed, list): raise ValueError("Parsed JSON is not a list") return parsed def infer_subclaims( medical_text: str, client: OpenAI, model: str = "gpt-5-mini", retries: int = 1, ) -> List[str]: if not medical_text or medical_text.strip() == "": return [] prompt = extraction_prompt(medical_text) try: response = client.chat.completions.create( model=model, messages=[ {"role": "system", "content": "Return only a valid JSON list of strings."}, {"role": "user", "content": prompt}, ], ) output_text = response.choices[0].message.content.strip() return _parse_json_list(output_text) except Exception as e: if retries > 0: time.sleep(1.5) return infer_subclaims( medical_text, client, model=model, retries=retries - 1, ) return [f"ERROR: {str(e)}"] # ----------------------------- # MAIN EXECUTION # ----------------------------- if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument( "--input_file", type=str, default="/home/mshahidul/readctrl/data/annotators_validate_data_(20_80)/combine/verified_combined_0-80.json", ) parser.add_argument( "--save_folder", type=str, default="/home/mshahidul/readctrl/data/extracting_subclaim", ) parser.add_argument("--model", type=str, default="gpt-5-mini") args = parser.parse_args() input_file = args.input_file save_folder = args.save_folder file_name = os.path.basename(input_file).split(".json")[0] output_file = os.path.join(save_folder, f"extracted_subclaims_{file_name}.json") Path(save_folder).mkdir(parents=True, exist_ok=True) client = _load_openai_client() with open(input_file, "r") as f: data = json.load(f) result = [] if os.path.exists(output_file): with open(output_file, "r") as f: result = json.load(f) def _item_key(obj: dict) -> str: if obj.get("index") is not None: return str(obj.get("index")) if obj.get("id") is not None: return str(obj.get("id")) if obj.get("doc_id") is not None and obj.get("label") is not None: return f"{obj.get('doc_id')}_{obj.get('label')}" return str(obj.get("doc_id") or obj.get("label") or "") processed_data = {_item_key(item): item for item in result} for item in tqdm.tqdm(data): item_id = _item_key(item) existing_entry = processed_data.get(item_id) # 1. Process Fulltext if not existing_entry or not isinstance(existing_entry.get("fulltext_subclaims"), list): f_sub = infer_subclaims( item.get("fulltext", ""), client, model=args.model, retries=2, ) else: f_sub = existing_entry["fulltext_subclaims"] # 2. Process Summary if not existing_entry or not isinstance(existing_entry.get("summary_subclaims"), list): s_sub = infer_subclaims( item.get("summary", ""), client, model=args.model, retries=1, ) else: s_sub = existing_entry["summary_subclaims"] # 3. Process Generated Texts (diff_label_texts) diff_label_texts = item.get("diff_label_texts", "") if isinstance(diff_label_texts, dict): diff_label_subclaims = existing_entry.get("diff_label_subclaims", {}) if existing_entry else {} for label, text in diff_label_texts.items(): if label not in diff_label_subclaims or not isinstance(diff_label_subclaims[label], list): diff_label_subclaims[label] = infer_subclaims( text, client, model=args.model, retries=1, ) else: if not existing_entry or not isinstance(existing_entry.get("diff_label_subclaims"), list): diff_label_subclaims = infer_subclaims( diff_label_texts, client, model=args.model, retries=1, ) else: diff_label_subclaims = existing_entry["diff_label_subclaims"] # 4. Save new_entry = { "doc_id": item.get("doc_id"), "label": item.get("label"), "fulltext": item.get("fulltext", ""), "fulltext_subclaims": f_sub, "summary": item.get("summary", ""), "summary_subclaims": s_sub, "diff_label_texts": diff_label_texts, "diff_label_subclaims": diff_label_subclaims, } processed_data[item_id] = new_entry if len(processed_data) % 10 == 0: with open(output_file, "w") as f: json.dump(list(processed_data.values()), f, indent=4, ensure_ascii=False) with open(output_file, "w") as f: json.dump(list(processed_data.values()), f, indent=4, ensure_ascii=False) print(f"Extraction completed. File saved at: {output_file}")