TL;DR
This paper introduces PERSPECTIVESUMM, a new benchmark and suite of metrics designed to evaluate how fairly Large Language Models (LLMs) represent diverse viewpoints in abstractive summarization. The authors discover that both LLMs and human-written references frequently underrepresent minority perspectives, but fairness can be improved through strategic prompt engineering and hyperparameter tuning.
What Problem Does This Paper Address?
When we ask an AI to summarize a heated debate or a product with mixed reviews, we expect a balanced overview. However, summarization has traditionally been treated as a compression task—reducing length while maintaining "salience." The problem is that "salience" is subjective. If a summary of 100 reviews (80 positive, 20 negative) only mentions the praise, it is technically accurate to the majority but unfair to the minority perspective.
Current evaluation metrics like ROUGE or BERTScore are blind to this. They measure how similar a summary is to a human reference, but if the human reference itself is biased, the metric reinforces that bias. This paper tackles the lack of formal definitions, datasets, and metrics for fair abstractive summarization, ensuring that summaries don't systematically silence specific social or demographic groups.
Methods
The authors define fairness as a distribution alignment problem: the proportion of perspectives in the summary should reflect the proportion in the source text (Ratio Fairness) or represent all groups equally (Equal Fairness).
To measure this without needing "perfect" human references, they propose a two-stage framework:
- Produce Value Distribution: They map summary content back to source groups using three methods: N-gram matching (exact word overlap), BERTScore (sentence-level semantic similarity), and BARTScore (summary-level entailment).
- Compute Metrics: They introduce four new metrics:
- Binary Unfair Rate (BUR): A "hard" metric that flags a summary as unfair if any group falls below a representation threshold ($\tau$).
- Unfair Error Rate (UER): A "soft" metric measuring the total distance between the source and summary distributions.
- Area Under Curve (AUC): Evaluates fairness across all possible thresholds.
- Second-Order Fairness (SOF): Measures the variance of unfairness across different groups to see if the model is biased toward one specific group.
Datasets
The authors curated PERSPECTIVESUMM, a benchmark unifying six existing datasets across diverse domains and social attributes:
- Claritin (Healthcare): Gender-based perspectives from Twitter.
- US Election (Politics): Political party affiliation from Twitter.
- Yelp (Restaurant): Sentiment-based reviews.
- Amazon (Product): Star ratings.
- Supreme Court & IQ2 (Law/Policy): Speaker-based perspectives from transcripts and debates.
Key Insights and Ideas
- Humans are Biased Too: Surprisingly, human-written reference summaries often performed worse on fairness metrics than LLMs. This suggests that humans naturally inject their own biases or focus only on the "majority" when writing summaries.
- LLaMA vs. GPT: In the battle of the models, the LLaMA 2 family (especially the 13B and 70B versions) generally demonstrated better fairness than the GPT family across several datasets.
- The Length Trap: There is a "Goldilocks zone" for summary length. Summaries that are too short lack the space to include minority views, while summaries that are too long can become repetitive or lose coherence. Medium-length summaries (around 3 sentences in their tests) were the most fair.
- Temperature Matters: Increasing the decoding temperature (making the model more "creative" or stochastic) actually improved fairness. By sampling less likely tokens, the model was more likely to include diverse perspectives that it might otherwise skip in a greedy search.
Key Results
The evaluation of nine major LLMs (including GPT-4, PaLM 2, and Claude) revealed widespread fairness issues. For instance, GPT-4 reached a BUR of 74.27% on Amazon reviews, meaning nearly 3 out of 4 summaries were biased toward specific rating groups.
However, the authors found three simple ways to fight back:
- Fairness Instructions: Simply adding a definition of fairness to the prompt reduced the BUR for GPT-3.5 from 51.11% to 47.48%.
- Hyperparameter Tuning: Raising the temperature from 0 to 1 consistently lowered both BUR and UER.
- Length Control: Prompting for a specific number of sentences helped balance the representation of different groups.
Limitations and Future Work
The authors acknowledge several constraints:
- Dataset Bias: The study focuses on model fairness, assuming the source datasets are the "ground truth" for distribution, though the source data itself may contain societal biases.
- Attribute Simplification: For experimental control, some attributes like gender were treated as binary, which does not reflect the full spectrum of identity.
- Granularity: Neural metrics like BARTScore rely on the accuracy of the underlying backbone model, which can introduce its own errors.
Future work aims to fine-tune models specifically for fair summarization and extend these metrics to other generative tasks like dialogue and open-ended storytelling.
Summary updated