EMNLP 2025 Main, "Probability Distribution Collapse: A Critical Bottleneck to Compact Unsupervised Neural Grammar Induction"

On this page

Video will be uploaded after conference!

Summary & Contribution

  • We introduce probability distribution collapse as a bottleneck to inducing more compact grammars in an unsupervised setting.

  • We investigate its causes within neural parameterization and propose a simple, yet effective solution, termed collapse-relaxing neural parameterization.

  • We provide extensive validation of our approach, demonstrating improvements in both upper bound and practical accuracy, as well as grammar compactness, on constituent parsing tasks for English and multilingual benchmarks.