division.py

import os
import json
from collections import defaultdict
from transformers import LlamaTokenizer
from tqdm import tqdm

input_json_file = '/Users/liyu/PycharmProjects/CoTKR/updated.json'
output_dir = '/Users/liyu/PycharmProjects/CoTKR/output'
os.makedirs(output_dir, exist_ok=True)

model_path = '/Users/liyu/PycharmProjects/CoTKR/Llama-2-7b-hf'

tokenizer = LlamaTokenizer.from_pretrained(model_path)

with open(input_json_file, 'r') as f:
    triples = json.load(f)

grouped_entity_triples = defaultdict(list)
for triple in triples:
    subj, relation, obj = triple
    grouped_entity_triples[subj].append(f"{relation}: {obj}")

# 根据Tokenizer的长度限制创建段落
def create_paragraphs(subj, relations, tokenizer, max_tokens=512):
    paragraphs = []
    current_chunk = []
    current_tokens = 0

    for relation in relations:

        tokens_in_relation = len(tokenizer.encode(relation, add_special_tokens=False))


        if current_tokens + tokens_in_relation + 1 > max_tokens:  # +1考虑分隔符或空格
            paragraphs.append({
                "title": f"{subj}",
                "contents": ", ".join(current_chunk).strip()
            })
            current_chunk = [relation]
            current_tokens = tokens_in_relation
        else:
            current_chunk.append(relation)
            current_tokens += tokens_in_relation + 1  # +1考虑分隔符或空格

    if current_chunk:
        paragraphs.append({
            "title": f"{subj}",
            "contents": ", ".join(current_chunk).strip()
        })

    return paragraphs

all_paragraphs = []
for subj, relations in tqdm(grouped_entity_triples.items(), desc="Processing subjects", unit="subject"):
    paragraphs = create_paragraphs(subj, relations, tokenizer)
    all_paragraphs.extend(paragraphs)

output_file = os.path.join(output_dir, 'new_triple_processed.json')
with open(output_file, 'w') as out_f:
    json.dump(all_paragraphs, out_f, ensure_ascii=False, indent=4)

# print(f"Processed paragraphs saved to {output_file}")