: Evaluating faithfulness and content selection in book-length summarization

This repo hosts FABLES (Faithfulness Annotations for Book-Length Summarization), a dataset of model-generated summaries with atomic claims annotated for faithfulness. Labels are supported by evidence and the annotator's reasoning, while claim sets include a general comment on the quality of the summary.

Authors: Yekyung Kim, Yapei Chang, Marzena Karpinska, Aparna Garimella, Varun Manjunatha, Kyle Lo, Tanya Goyal, Mohit Iyyer

Introduction

While long-context large language models (LLMs) can technically summarize book-length documents (> 100K tokens), the length and complexity of the documents have so far prohibited evaluations of input-dependent aspects like faithfulness. In this paper, we conduct the first large-scale human evaluation of faithfulness and content selection on LLM-generated summaries of fictional books. Our study mitigates the issue of data contamination by focusing on summaries of books published in 2023 or 2024, and we hire annotators who have fully read each book prior to the annotation task to minimize cost and cognitive burden. In total, we collect annotations on 3,158 claims made in LLM-generated summaries of 26 books at a cost of $5.2K USD, which allows us to rank LLM summarizers based on faithfulness: CLAUDE-3-OPUS significantly outperforms all closed-source LLMs, while the open-source MIXTRAL is on par with GPT-3.5-TURBO. An analysis of the annotations reveals that most unfaithful claims relate to events and character states, and they generally require indirect reasoning over the narrative to invalidate. While LLM-based auto-raters have proven reliable for factuality and coherence in other settings, we implement several LLM raters of faithfulness and find that none correlates strongly with human annotations, especially with regard to detecting unfaithful claims. Our experiments suggest that detecting unfaithful claims is an important future direction not only for summarization evaluation but also as a testbed for long-context understanding. Finally, we move beyond faithfulness by exploring content selection errors in book-length summarization: we develop a typology of omission errors related to crucial narrative elements and also identify a systematic over-emphasis on events occurring towards the end of the book. We will release our annotations to spur further research on the evaluation of book-length summarization.

Data

FABLES is a dataset created from book-length summaries of narrative books published in 2023-2024. The data contains:

🪄 Book Title -- Title of the book used for summarization. Each title is a dict contaitning data related to the book.

🪄 Model Name -- Name of the model that summariezed the book: MIXTRAL, GPT-3.5-TURBO, GPT-4, GPT-4-TURBO, or CLAUDE-3-OPUS. A dict containing all summary data for that model.

🪄 General Comment -- (str) Overview of issues with the set of claims, including omissions of important information, chronology, salience (mention of unimportant events/details), factuality, and others like repetitiveness or vagueness, with each summary receiving a comment.

🪄 Summary -- (str) Entire book summarized by one of five models: Mixtral, GPT-3.5-Turbo, GPT-4, GPT-4-Turbo, and Claude-3-Opus, using the hierarchical merging method described in Chang et al.. More specifically, for each step in hierarchical merging we used the prompts listed here.

🪄 Claims -- (dict) Numbered dictionary of decontextualized claims from each summary extracted using GPT-4.

🪄 Label -- (str) Faithfulness label assigned by native English-speaking annotators based on their reading. Labels are "Yes" (accurate reflection), "No" (misrepresentation), "PartialSupport" (partially corroborated), or "Inapplicable" (indeterminable), applied to each claim.

🪄 Reason -- (list) List of reasons provided by annotator' as a justification for each faithfulness label (provided for some claims).

🪄 Evidence -- (list) List of evidence, source text quotations, provided by annotators to support their labels (available for most claims).

FALBES data is stored in a JSON file, formatted as follows:

* `book title`: (str) 
  * `model name`: (str) 
    *  `summary`: (str) 
    *  `general_comment`: (str)
    *  `claims`: (dict)
       * `claim`: (str)
       * `label`: (str)
       * `evidence`: (list of str)
       * `reason`: (list of str)

The book title and model name are placeholders for their actual values.
In the claims dictionary, each claim is indexed numerically as a key, with each key pointing to a dictionary containing the fields claim, label, evidence, and reason.

Example entry:

{
  "Wildfire": {
    "CLAUDE-3-OPUS": {
      "general_comment": "This summary is largely chronological (though Aurora's decision to not attend her father's wedding comes after Russ' father's visit) and hits the main thematic elements of the text, though it disproportionately addresses the epilogue over other portions of the text. The claims are factual, however.",
      "summary": "Wildfire, the first book in the Icebreaker series, follows the love story of Aurora Roberts and Russ Callaghan, two college students who meet while working as counselors at Honey Acres, a sleepaway summer camp in California. Aurora, the daughter of a famous Formula 1 team owner, struggles with her distant father's lack of attention and affection, often acting out to gain his notice but facing continual disappointment. Russ, a reserved hockey player at UC Maple Hills, deals with the shame and embarrassment of his father's gambling addiction, which has strained their family relationships.(...)",
      "claims": {
        "0": {
          "claim": "The book 'Wildfire' is the first in the Icebreaker series.",
          "label": "No",
          "evidence": ["n/a"],
          "reason": ["Not quoted in the book, but this is the second in the series, after 'Icebreaker'."]
        }
      }
    }
  }
}

If you would like to access the book content dataset, please email to Yekyung with your affiliation and the purpose of your work.

Corpus Statistics

	Books (Documents)	Books (Summaries)	Books (Claims)	Annotations (Reasons)	Annotations (Evidence)	Annotations (Comments)
	(n=26)	(n=130)	(n=3,158)	(n=1,513)	(n=3,051)	(n=130)
Mean	121,467	594.3	19.8	37.6	194.7	155
St. dev.	35,836	119.5	6.4	33.4	218.5	148.4
Max	243,965	798	58	281	2,435	823
Min	49,762	172	6	2	5	6

Number of tokens across books and annotations (v1); based on tiktoken tokenizer.

Please note that the full source texts, being copyrighted, are not included in our release. However, we provide a list of the books for purchase, facilitating further investigation into summary accuracy and claim verification.

Disclaimer

The fictional narratives included in this dataset explore a range of themes, some of which may be sensitive, including mental health struggles such as depression, murder, and suicide.

Citation Information

If you use this dataset, please cite it as follows:

@misc{fables-2024-kim-et-al,
      title={FABLES: Evaluating faithfulness and content selection in book-length summarization}, 
      author={Yekyung Kim and Yapei Chang and Marzena Karpinska and Aparna Garimella and Varun Manjunatha and Kyle Lo and Tanya Goyal and Mohit Iyyer},
      year={2024},
      eprint={2404.01261},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

mungg/FABLES