Introduction
Relation Extraction (RE) is a critical task in Natural Language Processing, focused on identifying structured triples—such as (Subject, Relation, Object)—from unstructured text. These triples form the foundation of knowledge graphs and structured databases. However, the high cost of human annotation has led researchers to adopt Generative Data Augmentation (GDA), where Large Language Models (LLMs) are used to generate synthetic training data.
While GDA allows for rapid scaling, it introduces a significant challenge: "hallucinations." These occur when an LLM generates facts in the text that are not present in the provided labels. This paper investigates how different types of hallucinations affect RE model performance and proposes a robust detection method to mitigate their impact.
Categorizing Hallucinations
Not all hallucinations are equally detrimental. The authors distinguish between two primary types based on their relationship to the target schema:
- Relevant Hallucinations: These occur when the LLM adds a fact that belongs to a relation type the model is being trained to extract, but the fact is missing from the labels.
- Irrelevant Hallucinations: These involve facts about relations that fall outside the model's target schema (e.g., adding a person's height when the model only extracts birth dates).
The "Omission Trap" and Model Performance
The study's most significant finding is the "Omission Trap." When a training dataset contains relevant hallucinations, the model essentially learns that it is correct to ignore those specific facts because they are present in the text but absent from the labels. This training signal causes the model to become "lazy," leading to a massive drop in recall.
Conversely, RE models appear remarkably robust to irrelevant hallucinations. Because these facts fall outside the target schema, the model treats them as harmless noise, resulting in negligible performance impact.
As shown in the experimental results, relevant hallucinations caused recall to drop by 19.1% at the sentence level and a staggering 39.2% at the document level. Interestingly, precision often increased in these scenarios; models became extremely conservative, only extracting relations they were highly certain of, thereby avoiding false positives at the cost of missing many valid triples.
Detection and Mitigation
To address the toxicity of relevant hallucinations, the authors evaluated two automated detection strategies:
- NER-based Detection: This method uses Named Entity Recognition to check if entities in the text match the labels. However, it often fails due to synonyms or variations in phrasing (e.g., "Alan Bean" vs. "A. Bean").
- NLI-based Detection: By converting labeled triples into simple sentences and using a Natural Language Inference (NLI) model (specifically DeBERTa-v2-xlarge-mnli), the system checks if the generated text is logically entailed by the labels.
The NLI approach proved far superior, achieving an F1-score of 92.15% in identifying hallucinated samples. This suggests that automated filtering can effectively clean synthetic datasets before they are used for training.
Conclusion and Future Work
This research demonstrates that the quality of synthetic data is not just about the accuracy of the text, but the completeness of the labels. Relevant hallucinations act as a "poison" that specifically degrades model recall. While NLI-based detection offers a powerful solution for sentence-level data, scaling these methods to complex, multi-sentence documents remains a challenge. Future work will focus on integrating these detection mechanisms directly into GDA pipelines to enable the creation of high-fidelity synthetic datasets at scale.
Summary updated