Latent Tree Learning with Ordered Neurons: What Parses Does It Produce?

TL;DR

This paper investigates the parsing behavior of the ON-LSTM latent tree model, revealing its consistency and limitations in syntactic structure learning without supervision, and suggests potential improvements through alternative training tasks.

Contribution

The study replicates and analyzes ON-LSTM's parsing behavior, providing insights into its consistency and limitations, and proposes directions for enhancing unsupervised parsing models.

Findings

The model shows consistent parsing behavior across different restarts.

It struggles with complex noun phrase internal structures.

It tends to overestimate split point heights before verbs.

Abstract

Recent latent tree learning models can learn constituency parsing without any exposure to human-annotated tree structures. One such model is ON-LSTM (Shen et al., 2019), which is trained on language modelling and has near-state-of-the-art performance on unsupervised parsing. In order to better understand the performance and consistency of the model as well as how the parses it generates are different from gold-standard PTB parses, we replicate the model with different restarts and examine their parses. We find that (1) the model has reasonably consistent parsing behaviors across different restarts, (2) the model struggles with the internal structures of complex noun phrases, (3) the model has a tendency to overestimate the height of the split points right before verbs. We speculate that both problems could potentially be solved by adopting a different training task other than unidirectional language modelling.