TL;DR
This paper tackles the challenge of ensuring generated text accurately reflects its input data, a crucial aspect of data-to-text generation. The authors propose a set of detailed human annotation guidelines to address subtle "implicit phenomena" that often lead to factuality errors, providing a more robust framework for evaluating these systems.
What Problem Does This Paper Address?
Modern data-to-text generation systems, trained on vast datasets, are increasingly expected to produce text that is not only fluent but also factually accurate with respect to the input data. However, traditional evaluation metrics like BLEU, which focus on word overlap, fail to capture subtle factual discrepancies. Neural models can easily introduce information not present in the input ("extra") or omit information that is present ("missing"). This problem is exacerbated by the fact that training datasets themselves can contain misalignments between data and text. The core challenge this paper addresses is how to reliably evaluate the factuality of generated text, particularly when the input data and the generated text involve implicit information, common-sense reasoning, or require external world knowledge.
Methods
The authors' primary contribution is the development of a set of detailed human annotation guidelines designed to improve inter-annotator agreement when assessing factuality in data-to-text generation. They arrived at these guidelines by manually annotating data-text pairs from the WebNLG and E2E datasets and identifying systematic sources of uncertainty. These uncertainties were categorized into six types:
- Complex Label: Labels that convey additional or redundant data.
- Bridging Anaphora: Textual references that require external world knowledge to resolve.
- Aggregation: Text that aggregates data based on the semantics of a relation (e.g., collective or distributive).
- Pragmatic Inference: Text that implies information through presuppositions, scalar implicatures, or connectives.
- Units and Rounding: Inferred units for measurements or approximate numerical realizations.
- World Knowledge, Redundant Data, Implied Data: Cases where input data is implied but not explicit, logically redundant data is omitted, or implicit information (like a person's profession) is made explicit.
These guidelines aim to standardize how annotators handle these implicit phenomena, moving beyond simple surface-level matching.
Datasets
The paper uses data from two primary datasets for their analysis and annotation:
- WebNLG: A dataset for data-to-text generation.
- E2E: Another dataset for end-to-end data-to-text generation.
The authors also created synthetic data by perturbing existing data-text pairs (e.g., adding or removing triplets, altering arguments) to further test the robustness of their evaluation approach. They manually annotated 200 samples from WebNLG, including both original and perturbed data.
Key Insights and Ideas
- Implicit phenomena are a major source of factuality errors: The paper highlights that a significant portion of factuality errors in data-to-text generation stems from subtle implicit information, common sense, or world knowledge that is not explicitly stated in the input data.
- Traditional metrics are insufficient: Metrics like BLEU are inadequate for capturing these nuanced factual errors, necessitating more sophisticated evaluation methods.
- Human annotation is complex and requires clear guidelines: Even experienced annotators struggle with consistency when dealing with implicit information, underscoring the need for well-defined guidelines.
- Systematic categories of uncertainty exist: The authors' categorization of uncertainty into six types provides a structured way to understand and address the challenges in factuality assessment.
- Real-world data and synthetic perturbations reveal similar issues: The study found that both original and synthetically perturbed data exhibit similar frequencies of uncertainty, suggesting these implicit phenomena are inherent to the task and data.
Key Results
The authors manually annotated 200 samples from WebNLG. They found that:
- Over half of the annotated samples exhibited uncertainties: This indicates that implicit phenomena are a widespread challenge in data-to-text factuality.
- "Pragmatic Inference" and "World Knowledge" are the most frequent uncertainty categories: These two categories accounted for a substantial portion of the identified issues.
- High rate of non-reliable factuality in original data: 20% of the original WebNLG instances were annotated as non-reliable (missing, extra, or perturbation).
- Synthetic data exacerbates unreliability: 95% of the 100 noisy (synthetic) samples were annotated as non-reliable, demonstrating the impact of data perturbations on factuality.
- Low alignment with automatic assessment tools: The paper notes low alignment between their manual annotations and existing automatic assessment tools, further emphasizing the need for improved evaluation methods.
Limitations and Future Work
Limitations:
- Manual annotation is labor-intensive: While their guidelines aim to improve consistency, the process of manual annotation remains time-consuming and requires expert annotators.
- Focus on specific datasets: The guidelines were developed and tested on WebNLG and E2E, and their generalizability to other data-to-text datasets might require further validation.
- Subjectivity in interpretation: Despite guidelines, some degree of subjective interpretation may still exist in human annotation, especially for highly nuanced implicit phenomena.
Future Work:
- Developing automatic metrics: The authors suggest that their guidelines can serve as a basis for developing more robust automatic factuality evaluation metrics that can better capture implicit phenomena.
- Broader dataset coverage: Applying and refining these guidelines across a wider range of data-to-text datasets.
- Investigating model behavior: Using these guidelines to analyze the specific ways in which data-to-text models fail to handle implicit information and to guide future model development.
Summary updated