TL;DR
This paper introduces "Label-in-a-Haystack Rectification" (LiaHR), a novel method that leverages Large Language Models (LLMs) to identify and correct subjective annotation errors in datasets. By prompting LLMs to perform the annotation task with examples already in the prompt, LiaHR cleverly uses the model's failure to perfectly copy labels as a signal for potential errors, enabling more robust and nuanced subjective data processing.
What Problem Does This Paper Address?
Modeling subjective tasks like emotion recognition or moral judgment is notoriously difficult due to inherent human variation in interpretation. Unlike objective tasks where disagreement often signals error, in subjective tasks, this variation can represent legitimate differences in perspective. Traditional methods struggle to distinguish between genuine subjectivity and actual annotation errors, leading to a loss of valuable signal or the inclusion of noisy data. This paper tackles the challenge of filtering out errors while preserving meaningful subjective diversity in annotations, aiming to improve the quality of datasets for downstream NLP tasks.
Methods
The authors propose two main approaches. First, a "Reasonableness Baseline" uses LLMs in a few-shot setting to predict whether a document-label pair is "reasonable," acting as a binary filter. Second, and more significantly, they introduce the "Label-in-a-Haystack Rectification" (LiaHR) framework. In LiaHR, the LLM is prompted with examples, including the query document and its associated label. Crucially, the LLM is instructed to perform the actual task (e.g., emotion recognition) rather than just copy the label. The core idea is that if the LLM's internal reasoning disagrees with the provided label, it will fail to perfectly copy it to its output. This failure is then used as a signal to flag the label for potential correction. LiaHR can then generate alternative labels, effectively rectifying unreasonable annotations.
Datasets
The authors evaluate their methods on several existing datasets:
- SemEval 2018 Task 1 E-c: A multi-label emotion recognition benchmark.
- MFRC (Moral Foundation Corpus): A multi-label dataset for moral foundation prediction.
- GoEmotions: A multi-label emotion recognition benchmark. For efficiency, they cluster its 27 emotions into seven categories.
- QueerReclaimLex: A single-label binary harm dataset used as a stress test for diversity and bias.
Key Insights and Ideas
- LLMs as "Reasonableness" Detectors: The failure of LLMs to perfectly copy labels in a "Label-in-a-Haystack" setting is a surprisingly effective signal for identifying potentially erroneous subjective annotations.
- Preserving Subjectivity: LiaHR aims to filter out errors without discarding legitimate variations in human interpretation, a critical distinction for subjective tasks.
- Implicit Bias Detection: The QueerReclaimLex experiments show LiaHR can identify and even counteract implicit biases in LLMs, demonstrating robustness.
- Aggregation Artifacts: The study suggests that aggregating labels can lead to higher rejection rates, potentially masking genuine diversity.
Key Results
- LiaHR and the reasonableness baseline demonstrate the ability to identify and correct subjective labels.
- More capable LLMs (e.g., GPT-3.5, GPT-4, Llama-3 70b) show better adherence to the desired proxy properties like Nonconformity, Rectification, and Noise Rejection.
- Human evaluations indicate a preference for LiaHR-generated labels over original gold labels when correction is performed.
- Training smaller models on LiaHR-corrected datasets leads to slightly better generalization performance compared to using original labels.
- BERT-based filtering baselines underperform LiaHR, especially when considering the need for full dataset training.
Limitations and Future Work
- Not an Oracle: The authors emphasize that their model does not provide ground truth and can still be biased.
- Potential for Misuse: Deliberately skewing prompts in LiaHR could be used to justify rejecting minority labels or valid perspectives.
- Bias in Model Decisions: While robust on QueerReclaimLex, the model showed worse performance on in-group data, indicating potential biases.
- Quantization Effects: The use of 4-bit quantized LLMs might affect performance, though they argue it's representative of realistic scenarios.
- Future Work: The authors suggest further exploration of capturing nuance and variability in subjective tasks, incorporating a broader pool of annotators, and applying LiaHR in sensitive settings with caution. They also plan to investigate the impact of prompt position and further analyze objective tasks.
Summary updated