Introduction
Generative Data Augmentation (GDA) has emerged as a popular technique to scale Relation Extraction (RE) datasets by using Large Language Models (LLMs) to generate synthetic text-triple pairs. However, this "free" data often comes with a hidden cost: hallucinations. A recent study presented at KBC-LM’24 by researchers from the Karlsruhe Institute of Technology and TU Dresden investigates how these hallucinations—specifically facts present in text but missing from labels—impact model performance.
Defining the Hallucination Taxonomy
In the context of RE, hallucinations occur when the generated text contains information not reflected in the associated structured triples. The researchers categorize these into two types:
- Relevant Hallucinations: The text contains a fact that falls within the model's target schema (e.g., a birth date for a model trained to extract birth dates) but is missing from the training labels.
- Irrelevant Hallucinations: The text contains extra information outside the model's scope (e.g., an occupation for a birth date extraction model).
Experimental Setup and Methodology
To quantify the damage, the authors conducted experiments at both document and sentence levels. For document-level analysis, they leveraged the discrepancy between DocRED (a dataset known for incomplete annotations, effectively acting as hallucinated data) and Re-DocRED (a cleaned version with more complete labels).
At the sentence level, they used WebNLG and intentionally injected hallucinations using Llama 2 to observe the effects on state-of-the-art models like DREEAM and PFN.
Results: The "Lazy Model" Effect
The study's most significant finding is the "Lazy Model" effect caused by relevant hallucinations. When a model is repeatedly shown facts in text that should be extracted but are not labeled in the training data, it effectively learns that omitting these facts is correct behavior. This leads to a dramatic collapse in recall, with reductions ranging from 19.1% to 39.2%.
Conversely, irrelevant hallucinations have a negligible impact. Models appear robust enough to ignore noise that does not match their target schema, suggesting that the precision of the target schema itself acts as a filter. Interestingly, precision sometimes increased during these tests as models became extremely conservative, only extracting the most obvious triples.
Automated Detection via NLI
Recognizing the need for cleaner synthetic data, the authors evaluated automated detection methods. They found that Natural Language Inference (NLI) models significantly outperform traditional Named Entity Recognition (NER) approaches. By treating the triples as a premise and the generated text as a hypothesis, NLI models like DeBERTa-v2-xlarge can detect hallucinations with an F1-score of 92.15%.
Conclusion
While GDA offers a path to larger datasets, the quality of synthetic labels is paramount. Relevant hallucinations act as negative signals that degrade a model's ability to extract known relations. By integrating NLI-based filtering into the data generation pipeline, researchers can mitigate these effects and build more reliable Relation Extraction systems.
Summary updated