File size: 10,510 Bytes
633a5b1 | 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 | #!/usr/bin/env python3
"""
ZeroCoT Baseline: Zero-shot Chain-of-Thought Entity Alignment
=============================================================
Paper Table II: ZeroCoT* achieves Hits@1=0.639 on DBP15K without EvoRM,
and 0.711 (+11.3%) with EvoRM.
This baseline uses a simple zero-shot CoT prompt: "Let's think step by step"
followed by the entity matching decision. No reasoning framework, no rules.
Supports:
- Standalone mode (baseline)
- +EvoRM mode (plug-and-play enhancement)
"""
import os
import queue
import threading
import sys
from tqdm import tqdm
from openai import OpenAI
from collections import defaultdict
import json
sys.path.append('/root/autodl-tmp/AdaCoAgentEA')
from ThreadPoolExecutor import ThreadPoolExecutor
import tokens_cal
def load_entity_names(file_path):
entity_names = {}
with open(file_path, 'r', encoding='utf-8') as f:
for line in f:
parts = line.strip().split('\t')
if len(parts) == 2:
entity_names[int(parts[0])] = parts[1]
return entity_names
def load_triples(file_path):
triples = []
with open(file_path, 'r', encoding='utf-8') as f:
for line in f:
parts = line.strip().split('\t')
triples.append([int(x) for x in parts[:3]])
return triples
def get_entity_context(entity_id, entity_names, triples, rel_names, n=3):
relations = []
for h, r, t in triples:
if h == entity_id:
rel_str = rel_names.get(r, str(r))
tail_str = entity_names.get(t, str(t))
relations.append(f"- Has relation '{rel_str}' with {tail_str}")
elif t == entity_id:
rel_str = rel_names.get(r, str(r))
head_str = entity_names.get(h, str(h))
relations.append(f"- Is {rel_str} of {head_str}")
if len(relations) >= n:
break
context = f"Entity Name: {entity_names.get(entity_id, 'Unknown')}\n"
context += "Relationships:\n" + "\n".join(relations[:n])
return context
def group_candidates(input_file):
groups = defaultdict(list)
with open(input_file, 'r', encoding='utf-8') as f:
for line in f:
parts = line.strip().split('\t')
if len(parts) == 2:
try:
e1, e2 = int(parts[0]), int(parts[1])
groups[e1].append(e2)
except ValueError:
continue
return groups
def deduplicate_output_file(file_path):
if not os.path.exists(file_path):
return
unique_pairs = set()
with open(file_path, 'r', encoding='utf-8') as f:
for line in f:
line = line.strip()
if not line:
continue
parts = line.split('\t')
if len(parts) == 2 and parts[0] and parts[1]:
try:
e1, e2 = int(parts[0]), int(parts[1])
unique_pairs.add((e1, e2))
except ValueError:
continue
with open(file_path, 'w', encoding='utf-8') as f:
for e1, e2 in sorted(unique_pairs):
f.write(f"{e1}\t{e2}\n")
print(f"Deduplicated: {len(unique_pairs)} unique pairs")
# Zero-shot CoT prompt
ZERO_COT_PROMPT = """You are a knowledge graph entity alignment expert. Determine if two entities represent the same real-world object.
Think step by step:
1. Compare the entity names.
2. Compare the relationships of each entity.
3. Consider if the relationships describe the same real-world connections.
4. Make a decision.
Entity A (ID: {kg1_id}):
{context_a}
Entity B (ID: {kg2_id}):
{context_b}
Let's think step by step. After your analysis, output ONLY:
- The entity ID of the matching candidate (e.g., "{kg2_id}") if they match
- "No" if they do not match
Your response:"""
def align_entities(data_dir, from_m3=False, ablation_config=None,
no_optimization_tool=False, use_evorm=False,
evorm_ablation_mode=None):
"""
ZeroCoT Entity Alignment.
Args:
data_dir: Path to dataset directory
use_evorm: If True, wrap with EvoRM plugin
evorm_ablation_mode: EvoRM ablation mode (if use_evorm=True)
Returns:
List of aligned pairs [(kg1_id, kg2_id), ...]
"""
client = OpenAI(
base_url="https://hk.xty.app/v1",
api_key="sk-7a7Ev4VcVyysPLT5hqtqIVD6PybzJ1ZlEIVZddIR3NtZvPgK",
)
# EvoRM plugin
evorm_plugin = None
if use_evorm:
from evorm_plugin import EvoRMPlugin
evorm_plugin = EvoRMPlugin(
client=client,
ablation_mode=evorm_ablation_mode,
persistence_dir=os.path.join(data_dir, 'evorm_state_zero_cot'),
)
print(f"ZeroCoT + EvoRM enabled (ablation={evorm_ablation_mode})")
ent_names_1 = load_entity_names(os.path.join(data_dir, 'ent_ids_1'))
ent_names_2 = load_entity_names(os.path.join(data_dir, 'ent_ids_2'))
rel_names_1 = load_entity_names(os.path.join(data_dir, 'rel_ids_1'))
rel_names_2 = load_entity_names(os.path.join(data_dir, 'rel_ids_2'))
triples_1 = load_triples(os.path.join(data_dir, 'triples_1'))
triples_2 = load_triples(os.path.join(data_dir, 'triples_2'))
LLM1_PRIVATE_MESSAGE_POOL = {
'important_entities': os.path.join(data_dir, "message_pool", "important_entities.txt"),
'ucon_similarity_results': os.path.join(data_dir, "message_pool", "ucon_similarity_results.txt"),
'aligned_entities': os.path.join(data_dir, "message_pool", "aligned_entities.txt"),
}
input_file = LLM1_PRIVATE_MESSAGE_POOL['ucon_similarity_results'] if from_m3 else LLM1_PRIVATE_MESSAGE_POOL['important_entities']
output_file = LLM1_PRIVATE_MESSAGE_POOL['aligned_entities']
candidate_groups = group_candidates(input_file)
aligned_pairs = []
lock = threading.Lock()
executor = ThreadPoolExecutor(max_workers=5)
result_queue = queue.Queue()
def zero_cot_task(kg1_entity, kg2_candidates):
try:
context_a = get_entity_context(kg1_entity, ent_names_1, triples_1, rel_names_1)
for kg2_entity in kg2_candidates:
context_b = get_entity_context(kg2_entity, ent_names_2, triples_2, rel_names_2)
# Build prompt
prompt = ZERO_COT_PROMPT.format(
kg1_id=kg1_entity, kg2_id=kg2_entity,
context_a=context_a, context_b=context_b,
)
if evorm_plugin is not None:
# Build context dicts
es_ctx = {'entity_name': ent_names_1.get(kg1_entity, '')}
et_ctx = {'entity_name': ent_names_2.get(kg2_entity, '')}
for h, r, t in triples_1:
if h == kg1_entity:
rel = rel_names_1.get(r, str(r))
neighbor = ent_names_1.get(t, str(t))
key = f"neighbors_{rel}"
es_ctx.setdefault(key, set()).add(neighbor)
for h, r, t in triples_2:
if t == kg2_entity:
rel = rel_names_2.get(r, str(r))
neighbor = ent_names_2.get(h, str(h))
key = f"neighbors_{rel}"
et_ctx.setdefault(key, set()).add(neighbor)
# Stage 1: symbolic filtering
decision, triggered, candidates = evorm_plugin.stage1(
str(kg1_entity), str(kg2_entity), es_ctx, et_ctx)
if decision is not None:
if decision == 1:
with lock:
result_queue.put((kg1_entity, kg2_entity))
aligned_pairs.append((kg1_entity, kg2_entity))
continue
# Stage 2: LLM with evidence
result = evorm_plugin.stage2(
str(kg1_entity), str(kg2_entity), es_ctx, et_ctx,
triggered, candidates, prompt)
decision = result.get('decision')
if decision == 1:
with lock:
result_queue.put((kg1_entity, kg2_entity))
aligned_pairs.append((kg1_entity, kg2_entity))
evorm_plugin.record_trajectory(
str(kg1_entity), str(kg2_entity), es_ctx, et_ctx,
decision=decision or 0,
rationale=result.get('rationale', ''),
triggered_rules=triggered,
rule_feedback=result.get('rule_feedback', {}))
else:
# Baseline: direct LLM call
response = client.chat.completions.create(
model="gpt-3.5-turbo-1106",
messages=[{'role': 'user', 'content': prompt}],
temperature=0.1,
)
answer = response.choices[0].message.content.strip()
tokens_cal.update_add_var(response.usage.total_tokens)
if answer.lower() != "no" and str(kg2_entity) in answer:
with lock:
result_queue.put((kg1_entity, kg2_entity))
aligned_pairs.append((kg1_entity, kg2_entity))
break
except Exception as e:
print(f"Error processing entity {kg1_entity}: {str(e)}")
for kg1_entity_c, kg2_candidates_c in tqdm(candidate_groups.items(), desc="ZeroCoT"):
executor.submit(zero_cot_task, kg1_entity_c, kg2_candidates_c)
executor.shutdown(wait=True)
with open(output_file, 'a+', encoding='utf-8') as output_f:
while not result_queue.empty():
kg1_entity, kg2_id = result_queue.get()
output_f.write(f"{kg1_entity}\t{kg2_id}\n")
output_f.flush()
deduplicate_output_file(output_file)
if evorm_plugin:
stats = evorm_plugin.get_stats()
print(f"ZeroCoT+EvoRM stats: {json.dumps(stats, indent=2, default=str)}")
return aligned_pairs
if __name__ == "__main__":
data_dir = "/root/autodl-tmp/AdaCoAgentEA/data/icews_wiki"
print("ZeroCoT Baseline (no EvoRM):")
pairs = align_entities(data_dir, use_evorm=False)
print(f"Found {len(pairs)} aligned pairs.")
|