Resolving Gendered Ambiguous Pronouns with Gender-Fair Modeling Based on BERT Word Embeddings
Zhi Ling · 2023
Ambiguous anaphora resolution is a longstanding challenge in natural language understanding. This field possesses tremendous potential to improve the performance of other natural language processing (NLP) tasks like machine translation, sentiment analysis, paraphrase detection, summarization, etc. Previous studies have shown that current neural network based anaphora resolution systems all have the problems of gender bias, as these methods tend to behave better in male pronouns resolution than female. In this paper, we use the pre-trained BERT model combined with the natural language inference (NLI) architecture to transform the anaphora resolution task into a question answer (QA) extraction task, and employ a new data augmentation method to effectively overcome the gender bias problem. Without introducing any external data, we can achieve a single-model score (log loss) 0.20457 on the Gendered Ambiguous Pronouns (GAP) dataset released by the Google AI team through a Kaggle competition.