2019/06/09 by Rakesh Chada, Chada, Rakesh
Computer Science · #Natural Language Processing Techniques #Text Readability and Simplification #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1906.03695
The resolution of ambiguous pronouns is a longstanding challenge in Natural\nLanguage Understanding. Recent studies have suggested gender bias among\nstate-of-the-art coreference resolution systems. As an example, Google AI\nLanguage team recently released a gender-balanced dataset and showed that\nperformance of these coreference resolvers is significantly limited on the\ndataset. In this paper, we propose an extractive question answering (QA)\nformulation of pronoun resolution task that overcomes this limitation and shows\nmuch lower gender bias (0.99) on their dataset. This system uses fine-tuned\nrepresentations from the pre-trained BERT model and outperforms the existing\nbaseline by a significant margin (22.2% absolute improvement in F1 score)\nwithout using any hand-engineered features. This QA framework is equally\nperformant even without the knowledge of the candidate antecedents of the\npronoun. An ensemble of QA and BERT-based multiple choice and sequence\nclassification models further improves the F1 (23.3% absolute improvement upon\nthe baseline). This ensemble model was submitted to the shared task for the 1st\nACL workshop on Gender Bias for Natural Language Processing. It ranked 9th on\nthe final official leaderboard. Source code is available at\nhttps://github.com/rakeshchada/corefqa\n