Understanding Position Bias Effects on Fairness in Social Multi-Document Summarization

Olubusayo Olabisi, Ameeta AgrawalNAACL 2024

TL;DR

This paper investigates "position bias" in social media summarization, where models might favor information from earlier parts of the input. They found that while text quality remains consistent, ordering social media posts by dialect group significantly skews summaries, favoring the group that appears first. This highlights a critical fairness issue in multi-document summarization of diverse data.

What Problem Does This Paper Address?

Natural Language Processing (NLP) models are increasingly used to summarize vast amounts of text from diverse sources, including social media. While model quality (fluency, relevance) is crucial, ensuring fairness – that summaries accurately represent the viewpoints of different social groups – is equally important. A known issue in news summarization is "position bias" or "lead bias," where models tend to rely more on information presented at the beginning of a document. This paper explores whether this position bias exists and, more importantly, how it impacts the fairness of summaries generated from diverse social media data, specifically focusing on different linguistic communities.

Methods

The researchers conducted experiments using the DivSumm dataset, which contains tweets from three distinct linguistic communities: African-American English (AA), Hispanic-aligned Language (Hispanic), and White-aligned Language (White). They employed two main experimental setups:

  1. Shuffled: Input documents were presented in a random order, mimicking how they were originally presented in the dataset.
  2. Ordered: Input documents were grouped by their dialect group. They tested three specific orderings: White-aligned first (Ow), AA first (Oa), and Hispanic first (Oh).

They evaluated ten different summarization models (seven abstractive and three extractive) on these setups. To measure position bias, they analyzed the "token overlap" between the generated summaries and the input documents at different text positions. For fairness, they used semantic similarity to measure how well each dialect group was represented in the summary, calculating a "Fairness Gap" (ΔFair) as the difference between the maximum and minimum similarity scores across groups. Textual quality was assessed using standard metrics like ROUGE-L, BARTScore, BERTScore, and UniEval.

Datasets

The primary dataset used is DivSumm, an existing multi-document summarization dataset comprising English tweets from three diverse dialects: African-American English, Hispanic-aligned Language, and White-aligned Language. Each of the 25 topical sets in DivSumm contains 90 documents, evenly distributed (30 per dialect). The dataset also includes human-written extractive and abstractive summaries.

Key Insights and Ideas

  • Human Summaries are Fair: Interestingly, human-written reference summaries in the DivSumm dataset showed no position bias, suggesting annotators were not influenced by document order when creating summaries.
  • Shuffled Social Media is Fair: When social media posts were presented in a shuffled order, the system-generated summaries also exhibited no significant position bias and maintained a balanced representation across dialect groups.
  • Ordering Creates Bias: The core finding is that simply reordering the input documents by dialect group introduces a strong position bias. Models disproportionately favor information from the group that appears first in the ordered input.
  • Fairness Suffers, Quality Doesn't: While ordering severely impacted the fairness of summaries (leading to skewed representation), the textual quality (fluency, relevance) remained largely comparable across both ordered and shuffled inputs. This suggests that models can be "good" in terms of text quality but "bad" in terms of fairness due to input ordering.
  • Social Linguistic Similarity Matters: The paper notes that models tend to favor earlier text only when it exhibits social linguistic similarity, implying a nuanced form of position bias in diverse social data.

Key Results

  • Position Bias in Ordered Inputs: For ordered inputs, three out of seven abstractive models (BART, LED, PRIMERA) showed a strong position bias, with up to 3 times higher token overlap in the beginning of the input document from the favored group. Three other abstractive models showed weak bias, and three extractive models also showed bias.
  • Fairness Gap: In ordered scenarios, the fairness gap (ΔFair) varied significantly. For example, with Ow (White first), the gap was 0.14, while with Oa (AA first), it was 0.09. In contrast, the shuffled input consistently yielded the lowest fairness gap (0.04).
  • Textual Quality: Tables 2 and 6 show that quality metrics (ROUGE-L, BARTScore, BERTScore, UniEval) were comparable between ordered and shuffled inputs for both abstractive and extractive models, indicating that ordering did not negatively impact overall summary quality.

Limitations and Future Work

The authors acknowledge several limitations:

  • Offensive Language: Naturally occurring social media data may contain offensive language, which models might inadvertently reproduce in summaries.
  • Acronyms and Slangs: The use of acronyms and slang, which can have varied meanings across groups, is not explicitly accounted for.
  • Proxy Metrics for Fairness: The study uses semantic similarity as a proxy for fairness, acknowledging that the definition of fairness can be complex and vary for humans.
  • Racial Information: The dialectal information used for categorization should not be misconstrued as a direct representation of users' racial information.

Future work could explore more robust methods for mitigating this position bias, potentially through dataset augmentation (e.g., randomized shuffling) or by developing model architectures that are inherently less susceptible to input order when processing diverse social data. Further research into more nuanced fairness definitions and evaluation metrics for social media summarization is also warranted.

Summary updated

BibTeX
@inproceedings{olabisi2024understanding,
  author = {Olubusayo Olabisi and Ameeta Agrawal},
  title = {{Understanding Position Bias Effects on Fairness in Social Multi-Document Summarization}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure