Latent Tree Learning with Differentiable Parsers: Shift-Reduce Parsing and Chart Parsing

Abstract

Latent tree learning models represent sentences by composing their words according to an induced parse tree, all based on a downstream task. These models often outperform baselines which use (externally provided) syntax trees to drive the composition order. This work contributes (a) a new latent tree learning model based on shift-reduce parsing, with competitive downstream performance and non-trivial induced trees, and (b) an analysis of the trees learned by our shift-reduce model and by a chart-based model.

arXiv commentACL 2018 workshop on Relevance of Linguistic Structure in Neural Architectures for NLP

References in corpus (4)

• Adam: A Method for Stochastic Optimization
• Categorical Reparameterization with Gumbel-Softmax
• Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation
• Jointly Learning Sentence Embeddings and Syntax with Unsupervised Tree-LSTMs

Cited by in corpus (6)

• Jointly Learning Sentence Embeddings and Syntax with Unsupervised Tree-LSTMs
• Unsupervised Recurrent Neural Network Grammars
• Cooperative Learning of Disjoint Syntax and Semantics
• Analysis Methods in Neural Language Processing: A Survey
• Grammar-Based Grounded Lexicon Learning
• Dynamic Compositionality in Recursive Neural Networks with Structure-aware Tag Representations