Where Did That Come From? Sentence-Level Error-Tolerant Attribution

Ori Ernst, Aviv Slobodkin, Meng Cao, Sihui Wei, Jackie Ck CheungEMNLP 2025

Sure, here's a blog-style summary of the paper "Where Did That Come From? Sentence-Level Error-Tolerant Attribution":

TL;DR

This paper introduces "error-tolerant attribution," a new approach to understanding how generated text relates to its source. Instead of only attributing faithful content, it also accounts for hallucinations and errors, enabling better faithfulness evaluation, faster manual annotation, and more effective error correction.

What Problem Does This Paper Address?

Text generation models are increasingly used to summarize, answer questions, and explain information based on source documents. A critical concern is "faithfulness" – whether the generated text accurately reflects the source. When generated text contains errors or "hallucinations" (information not present in the source), it can mislead users and erode trust.

Traditional "attribution" tasks aim to identify which parts of the source support a generated output. However, existing methods often exclude unsupported or hallucinated content, leaving these errors unattributed. This limitation hinders the ability to fully verify content, locate errors, and fix them efficiently. Furthermore, existing attribution benchmarks often operate at the document level, making it time-consuming to pinpoint the exact source relevant to a specific generated sentence.

This paper addresses these limitations by proposing sentence-level error-tolerant attribution. This new definition extends attribution to include incorrect or hallucinated content, allowing for a more comprehensive understanding of the relationship between source and generated text.

Methods

The core contribution is the formalization of sentence-level error-tolerant attribution. This approach focuses on identifying the minimal set of source sentences that are most relevant for assessing the faithfulness of a single generated sentence. This set can include:

  • Evidence: Sentences directly supporting the generated text.
  • Contradiction: Sentences that explicitly disagree with the generated text.
  • Near Match: Sentences that are similar but not identical, possibly with minor modifications.
  • Expected Span: Sentences where the information should logically appear if it were present.
  • None: Indicating the generated sentence is fabricated or has no plausible basis in the source.

The authors also introduce a benchmark dataset and evaluate various models on this task. Their approach is "post-hoc," meaning attribution is inferred after the text has been generated, which is crucial for real-world applications where generated text is already fixed.

Datasets

The authors leverage the TofuEval benchmark, which comprises two summarization datasets: MediaSum (summarizing dialogues) and MeetingBank (summarizing meeting transcripts). They sampled 50 documents from each dataset and used summaries generated by six different LLM-based models. For their specific task, they manually annotated these summaries for faithfulness, error types, and detailed explanations. This resulted in a dataset split into development and test sets.

Key Insights and Ideas

  • Attribution for Errors is Key: The most significant insight is that attributing unfaithful content is as important, if not more so, than attributing faithful content. This allows for pinpointing errors and understanding why they occur.
  • Efficiency Gains: Sentence-level attribution significantly speeds up manual evaluation. Annotators can assess faithfulness much faster when provided with targeted source sentences.
  • Model Behavior Insights: The approach provides a lens into model behavior, revealing patterns in how models generate faithful vs. unfaithful content.
  • Hallucination Factors: The analysis of unfaithful outputs points to specific characteristics of source text that increase hallucination risk, such as sentences appearing later in the document or containing non-literal language.

Key Results

  • Faithfulness Evaluation: Most models showed improved performance on faithfulness evaluation when provided with attributed content, especially in an "attribution-only" mode.
  • Manual Annotation Speedup: Sentence-level attribution reduced manual annotation time by 30% in long-document settings, while slightly improving accuracy.
  • Error Correction: Using highlighted attributions led to better and more highly-ranked corrections of unfaithful sentences, particularly for smaller LLMs.
  • Hallucination Predictors: Unfaithful-linked sentences were more likely to occur in the last 10% of a document and to contain non-literal expressions.

Limitations and Future Work

  • Dataset Domain: The current conclusions are based on dialogue summarization datasets. The generalizability to other tasks like question answering remains uncertain.
  • Future Work: The authors suggest exploring additional features to combat hallucination, such as more sophisticated analysis of sentence position and non-literal language. They also highlight the potential for this framework to improve text generation systems by providing a diagnostic signal for model behavior.

Summary updated

BibTeX
@inproceedings{ernst2025where,
  author = {Ori Ernst and Aviv Slobodkin and Meng Cao and Sihui Wei and Jackie CK Cheung},
  title = {{Where Did That Come From? Sentence-Level Error-Tolerant Attribution}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure