IAAR-Shanghai/UHGEval-dataset

The full pipeline of creating UHGEval hallucination dataset

The full pipeline of creating UHGEval hallucination dataset

1. Collect the raw news

• Status: Full data; Avaliable.
• Data location: ./sources/xinhua/raw/
• Number: 75 txt files, 737,766 news in total
• Note: Those data are belong to Xinhua News Agency, and are only used for research purposes.

2. Preprocess the raw news

• Status: No data; Need to generate using the script.
• Script: ./sources/xinhua/preprocessor.py
• Data location: ./sources/xinhua/processed; Use the script to generate the data
• Number: Retained 25,005 news articles (constituting 3.39% of the raw news).
• Filtering settings:
  • Only includes news categories such as: '政治', '法律', '军事', '教育', '体育', '经济', '市场', '科学', '技术', '医疗', '卫生', '社会', '文化', '艺术', '娱乐', '天气', '环保', '灾害', '事故' ('Politics', 'Law', 'Military', 'Education', 'Sports', 'Economics', 'Market', 'Science', 'Technology', 'Medical', 'Health', 'Society', 'Culture', 'Art', 'Entertainment', 'Weather', 'Environmental Protection', 'Disaster', 'Accident').
  • The length of newsBeginning + newsRemainder is between [630, 870].
  • newsBeginning has [2, 5] sentences. Note: sentence-ending symbols include "。；：？！"
  • The length of newsBeginning is between [80, 120].

3. Generate candidates

• Status: No data; Need to generate using the script.
• Script: ./gen_candidates.py
• Data location: ./candidates/
• Number: Retained 17,503 news articles (constituting 70.00% of the preprocessed news).
• Filtering settings:
  • keywordPrecision is between (0, 1), generally should be between (0.2, 0.6).
  • candidateHallucinatedContinuation consists of only 1 sentence.
  • The length of candidateHallucinatedContinuation is between [20, 70].
  • appearedKeywords has at least 2 keywords.

4. Automatic labelling

• Status: Partial data as examples; Need to generate using the script.
• Script: ./gen_machine_annotations.py
• Data location: ./machine_annotations/keyword_hallucinated
• Note: Only articles labeled as having hallucinations were left for subsequent processing; those without hallucinations are located in ./machine_annotations/unhallucinated

5. Use Label Studio to enable human rechecking

Label Studio is a multi-type data labeling and annotation tool with standardized output format.

Relevant files can be found in ./label_studio_annotations/.

5.1 Prepare Label Studio Pre-annotations

• Status: No data; Need to generate using the script.
• Script: ./gen_pre_annotations.py
• Data location: ./label_studio_annotations/pre_annotations
• Reference: Import pre-annotated data into Label Studio

5.2 Setup labeling configuration and begin human rechecking

• Script: ./label_studio_annotations/Label Studio Config.xml
• An instruction for human annotators: ./label_studio_annotations/README.md
• An additional tutorial for human annotators: ./label_studio_annotations/新闻领域机器续写幻觉数据集标注培训.pptx
• Reference: Labeling configuration

5.3 Export Label Studio JSON annotations

• Status: Partial data as examples; Need to export using Label Studio.
• Data location: ./label_studio_annotations/after_annotations
• Reference: Export annotations and data from Label Studio

6. Get final hallucination dataset

• Status: Full data; Avaliable.
• Script: ./gen_hallucinations.py
• Data location: ./hallucinations/XinhuaHallucinations.json
• Number: Retained 5,141 news articles (constituting 29.37% of the candidates).