TL;DR
This paper tackles the issue of gender bias in coreference resolution by introducing a new dataset, GAP-Subjective, which augments the existing GAP dataset with subjective sentences. By using style transfer to convert objective sentences to subjective ones, they demonstrate that the presence of subjectivity significantly impacts how models perceive and exhibit gender bias, suggesting that current evaluations might be incomplete.
What Problem Does This Paper Address?
Coreference resolution is a fundamental NLP task that involves identifying all expressions in a text that refer to the same entity. Gender bias in NLP models is a well-documented problem, and the GAP dataset was created to evaluate and mitigate this bias in coreference resolution. However, the original GAP dataset primarily consists of objective sentences. Subjectivity, which involves expressing opinions, evaluations, and viewpoints, is pervasive in everyday language (news, journals, blogs). The authors argue that evaluating coreference models solely on objective text might not fully capture their behavior in real-world scenarios where subjectivity is prevalent. Therefore, they aim to address this gap by creating a new evaluation benchmark that includes subjective sentences, allowing for a more comprehensive analysis of gender bias.
Methods
The core of this paper is the creation of the GAP-Subjective dataset, which involves a multi-step process:
- Subjectivity Detection: They first identify objective sentences within the original GAP dataset. This is achieved by fine-tuning a RoBERTa model on the Wiki Neutrality Corpus (WNC), a dataset of objective and subjective sentence pairs.
- Style Transfer: Objective sentences are then transformed into their subjective counterparts using a sequence-to-sequence (Seq2Seq) model. This model is trained on the WNC dataset using a SentencePiece tokenizer and a BiLSTM architecture with a copy mechanism. The copy mechanism is crucial for preserving important entity-linked information, such as pronouns and entity names, which are vital for coreference resolution.
- Thresholding: To ensure the quality of the transferred sentences, they apply three thresholding techniques:
- Fluency: Perplexity is used to measure how natural the transferred sentences sound, discarding those with significant changes from the original.
- Content Preservation: Sentence-level GLEU scores are used to ensure that the core meaning of the original sentence is retained.
- Original Attribute: They filter out sentences that were already subjective before the transfer process.
- Offset Finding: After generating the subjective sentences, they recalculate the offsets for pronouns and entities to match the new sentence structure.
Datasets
- GAP Dataset: The original gender-balanced evaluation corpus for coreference resolution, consisting of 8,908 ambiguous pronoun-name pairs derived from Wikipedia. It primarily contains objective sentences.
- Wiki Neutrality Corpus (WNC): Used for training the subjectivity detection and style transfer models. It contains aligned pre- and post-neutralization sentences from Wikipedia revisions.
- GAP-Subjective Dataset: The newly introduced benchmark dataset. It is created by transforming objective sentences from the GAP dataset into subjective ones using style transfer. It retains the original GAP structure with added subjective variants.
Key Insights and Ideas
- Subjectivity Matters for Bias: The most striking insight is that the presence of subjectivity significantly alters how a BERT model exhibits gender bias. A model that prefers male entities on objective sentences shows a slight preference for female entities on subjective sentences. This suggests that bias evaluation is highly dependent on the linguistic style of the data.
- Style Transfer as a Tool for Data Augmentation: The paper demonstrates the effectiveness of using style transfer to create new, valuable evaluation data. This approach allows for controlled augmentation of existing datasets to explore specific linguistic phenomena like subjectivity.
- Preserving Coreference Information is Key: The use of a copy mechanism in the Seq2Seq model for style transfer is highlighted as crucial for maintaining the integrity of coreference resolution tasks, ensuring that entity and pronoun information is not lost during the transformation.
Key Results
- Subjectivity in GAP: The authors found that over 86% of sentences in the original GAP dataset are objective.
- Style Transfer Quality: Automated evaluations showed good fluency and content preservation in the transferred sentences. Human evaluation indicated that transferred sentences scored higher on subjectivity (average 1.21 points higher) with only a minor decrease in fluency (0.23 points).
- Benchmarking Performance: When benchmarking a BERT-based model:
- Overall F1: Decreased slightly from 0.796 (GAP) to 0.789 (GAP-Subjective).
- Male F1: Decreased by approximately 2% (0.802 to 0.786).
- Female F1: Increased slightly by approximately 0.3% (0.790 to 0.792).
- Gender Bias (B): Shifted from a slight male preference (0.984 on GAP) to a slight female preference (1.007 on GAP-Subjective).
Limitations and Future Work
- Limitations:
- The style transfer process, while effective, resulted in a slight decrease in overall predictive performance (F1 score) and a shift in gender bias.
- The authors acknowledge that the offset finding process was not perfect, with some instances not having correctly calculated offsets.
- Future Work:
- Increasing the coverage of objective-heavy datasets for other downstream tasks.
- Exploring the impact of other attributes (beyond subjectivity) on gender bias in coreference resolution.
- Further refining the style transfer and thresholding techniques.
Summary updated