Code and Data for "On Positional Bias of Faithfulness for Long-form Summarization"

By David Wan, Jesse Vig, Mohit Bansal, and Shafiq Joty.

1. Structure

• data/generation: Data and code for replicating data for generating summaries
• data/metric_benchmark: Data and code for replicating data for meta-evaluation
• src: Code for running the metric and different generation methods
• prompts: All prompts used for generation

2. Data Replication

This section describes how to replicate the datasets used for both the metric meta-evaluation and the summary generation tasks. All processed data and processing scripts are located in the data/ directory.

2.1 Metric Meta-Evaluation Benchmark

The benchmark data is used to evaluate the performance of faithfulness metrics. Each line in the jsonl files has the following structure:

• id: A unique identifier.
• documents: A list of source documents.
• document: The concatenated original document.
• summary: The system-generated summary to be evaluated.
• faithfulness: A binary label indicating faithfulness (1 for faithful, 0 for not).
• ranking: A list of document indices, ordered from most to least important.

DiverseSumm

1. Download the original annotations from the DiverseSumm repository.
2. Run the processing script:

   python data/metric_benchmark/process_diversesumm.py ${path_to_diversesumm_annotation_jsonl}
   

LongEval

1. Clone the LongEval repository to obtain the sentence-level faithfulness annotations.
2. Place the pubmed_annotations and squality_annotations directories into data/metric_benchmark/longeval/.
3. Download the pubmed_test.txt file from Cohan et al., 2018 and place it in the same directory.
4. Note: The script process_longeval.py contains hardcoded paths. You may need to adjust the paths on lines 13, 14, 28, 116, 119, and 122 to match your file locations.
5. Run the processing script:

   python data/metric_benchmark/process_longeval.py
   

Other Benchmarks (ChemSumm, ArXiv, etc.)

1. Download the compiled annotations from Infuse.
2. From the Infuse data, place DiverSumm.csv into data/metric_benchmark/diversumm/.
3. Download the ArXiv and PubMed test sets from Cohan et al., 2018 and place arxiv_test.txt into data/metric_benchmark/scientific_papers/.
4. Note: You may need to update the hardcoded paths in process_other.py (e.g., lines 6 and 72) if your file structure differs.
5. Run the processing script:

   python data/metric_benchmark/process_other.py
   

2.2 Generation Data

This data is used as input for the various summary generation methods. Each jsonl file contains lines with id and document (a list of documents).

• DiverseSumm: Use the data from the original authors and run:

  python data/generation/process_diversesumm.py
  

• Other Datasets: We use the datasets library. Please refer to the corresponding scripts in data/generation/ for more details.

3. Running the Code

All source code is located in the src/ directory. Prompts for all models and tasks are in prompts/.

3.1 Metric Meta-Evaluation

To evaluate generated summaries using an LLM-based metric (e.g., GPT-4o), use src/evaluate_summaries.py.

Example:

python src/evaluate_summaries.py \
  --model_name gpt4o \
  --data_file data/metric_benchmark/arxiv.jsonl \
  --output_file results/arxiv_eval.json \
  --document_merge_type max

3.2 Summary Generation

All prompts used for generation can be found in the prompts/ directory.

Standard & "Focus" Generation

Use src/generate_summaries.py. The "Focus" method uses a modified prompt (prompts/arxiv_prompt_top.txt) to guide the model.

python src/generate_summaries.py \
  --model_name gpt4o \
  --data_file data/generation/arxiv.jsonl \
  --prompt_file prompts/arxiv.txt \
  --output_file results/summaries_standard.json

Incremental Updating

python src/generate_summaries_incremental.py \
  --model_name gpt4o \
  --data_file data/generation/arxiv.jsonl \
  --prompt_file prompts/arxiv_incremental.txt \
  --original_prompt_file prompts/arxiv.txt \
  --output_file results/summaries_incremental.json

Hierarchical Merging

This is a two-step process.

Step 1: Generate summaries for each document individually.

python src/generate_summaries_individual.py \
  --model_name gpt4o \
  --data_file data/generation/arxiv.jsonl \
  --prompt_file prompts/arxiv.txt \
  --output_file results/summaries_individual.json

Step 2: Merge the individual summaries.

python src/generate_summaries_merge.py \
  --model_name gpt4o \
  --prompt_file prompts/arxiv_merge.txt \
  --data_file results/summaries_individual.json \
  --output_file results/summaries_hierarchical.json

Calibration

This process is similar to hierarchical merging but uses a different initial generation script.

Step 1: Generate summaries for each document with calibration.

python src/generate_summaries_calibration_initial.py \
  --model_name gpt4o \
  --data_file data/generation/arxiv.jsonl \
  --prompt_file prompts/arxiv.txt \
  --output_file results/summaries_calibration_individual.json

Step 2: Merge the individual summaries.

python src/generate_summaries_merge.py \
  --model_name gpt4o \
  --prompt_file prompts/arxiv_merge.txt \
  --data_file results/summaries_calibration_individual.json \
  --output_file results/summaries_calibration.json

4. Pre-computed Outputs

We provide all our generated outputs and metric scores in the following Google Drive folder:

Link to Pre-computed Outputs

• Metric Scores: These are JSON files containing sentence-level faithfulness scores. The structure is a list of lists: [num_examples, num_documents, num_sentences].
  • Scores from MiniCheck are floats; scores from GPT-4o are strings.
  • The _split variant contains scores for each document separately. The _full variant contains scores for the concatenated document.
• Generated Summaries: These are JSON files that mirror the input data format but include an additional generated_summary field. We also include the MiniCheck scores for all generated outputs.

5. Citation

If you find our work useful in your research, please cite our paper:

@inproceedings{wan-etal-2025-positional,
    title = "On Positional Bias of Faithfulness for Long-form Summarization",
    author = "Wan, David  and Vig, Jesse  and Bansal, Mohit  and Joty, Shafiq",
    editor = "Chiruzzo, Luis  and Ritter, Alan  and Wang, Lu",
    booktitle = "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
    month = apr,
    year = "2025",
    address = "Albuquerque, New Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.naacl-long.442/",
    doi = "10.18653/v1/2025.naacl-long.442",
    pages = "8791--8810",
    ISBN = "979-8-89176-189-6",
}