修复 MMKG 子图/QA 三元组解析并补全示例图片

MANIFEST.in

@@ -6,8 +6,8 @@ recursive-include dataflow *.sh
 recursive-include dataflow *.txt
 
 # 包含样例文件夹下所有类型的文件
-# 样例文件夹路径为dataflow/exaple
-recursive-include dataflow/example */
+# 样例文件夹路径为dataflow/example
+recursive-include dataflow/example *
 
 # 包含顶层文档文件
 include README.md

dataflow/example/MultimodalKGPipeline/input.json

@@ -1,6 +1,13 @@
 [
   {
     "raw_chunk": "Tesla unveiled the Cybertruck at a product event. Elon Musk appeared on stage during the launch presentation. The presentation focused on electric vehicle design and manufacturing.",
-    "img_dict": {}
+    "img_dict": {
+      "img_cybertruck": "../example_data/MultimodalKGPipeline/images/cyber.jpg",
+      "img_musk_stage": "../example_data/MultimodalKGPipeline/images/musk.jpg"
+    },
+    "vis_url": [
+      "../example_data/MultimodalKGPipeline/images/cyber.jpg",
+      "../example_data/MultimodalKGPipeline/images/musk.jpg"
+    ]
   }
 ]

dataflow/operators/multi_model_kg/filter/mmkg_visual_triple_subgraph_sampling.py

@@ -191,8 +191,8 @@ def _extract_vis_urls(self, entities, vis_triples, img_dict):
 
         for vt in vis_triples:
 
-            subj_match = re.search(r"<subj>\s*(.+?)\s*(?=<rel>)", vt)
-            obj_match = re.search(r"<obj>\s*(.+?)\s*$", vt)
+            subj_match = re.search(r"<subj>\s*(.+?)\s*(?=<obj>)", vt)
+            obj_match = re.search(r"<obj>\s*(.+?)\s*(?=<rel>)", vt)
 
             if not subj_match or not obj_match:
                 continue
@@ -283,7 +283,7 @@ def run(
 
             for vt in vis_triples:
 
-                subj_match = re.search(r"<subj>\s*(.+?)\s*(?=<rel>)", vt)
+                subj_match = re.search(r"<subj>\s*(.+?)\s*(?=<obj>)", vt)
 
                 if subj_match:
 

dataflow/operators/multi_model_kg/generate/mmkg_visual_triple_subgraph_qa_generator.py

@@ -1,4 +1,5 @@
 import json
+import re
 from typing import List, Dict, Any
 from tqdm import tqdm
 
@@ -115,15 +116,24 @@ def run(
         all_qas = []
 
         for _, row in tqdm(df.iterrows(), total=len(df), desc="Generating QA"):
-            # 构建 img_dict: key 从 vis_triple 中抽取图片ID, value 从 vis_url
+            # 构建 img_dict: key 从 vis_triple 中按首次出现顺序去重抽取图片ID, value 从 vis_url
+            # 注意: vis_url 在 step4 中已按首次出现顺序去重, 所以这里 img_id 也按首次出现顺序去重才能正确配对
             img_dict = {}
             vis_url_list = row.get(input_key, [])
             vis_triple_list = row.get("vis_triple", [])
 
-            # 提取图片ID对应 URL
-            for triple, url in zip(vis_triple_list, vis_url_list):
-                # triple 格式: "<subj> X <rel> <obj> img_ID"
-                img_id = triple.strip().split()[-1]
+            ordered_img_ids = []
+            seen_img_ids = set()
+            for triple in vis_triple_list:
+                m = re.search(r"<obj>\s*(.+?)\s*(?=<rel>)", triple)
+                if not m:
+                    continue
+                img_id = m.group(1).strip()
+                if img_id not in seen_img_ids:
+                    seen_img_ids.add(img_id)
+                    ordered_img_ids.append(img_id)
+
+            for img_id, url in zip(ordered_img_ids, vis_url_list):
                 img_dict[img_id] = url
 
             subgraph = row.get(input_key_meta, [])