Alert button

A Span-based Linearization for Constituent Trees

Apr 30, 2020
Yang Wei, Yuanbin Wu, Man Lan

Figure 1 for A Span-based Linearization for Constituent Trees
Figure 2 for A Span-based Linearization for Constituent Trees
Figure 3 for A Span-based Linearization for Constituent Trees
Figure 4 for A Span-based Linearization for Constituent Trees

Share this with someone who'll enjoy it:

We propose a novel linearization of a constituent tree, together with a new locally normalized model. For each split point in a sentence, our model computes the normalizer on all spans ending with that split point, and then predicts a tree span from them. Compared with global models, our model is fast and parallelizable. Different from previous local models, our linearization method is tied on the spans directly and considers more local features when performing span prediction, which is more interpretable and effective. Experiments on PTB (95.8 F1) and CTB (92.4 F1) show that our model significantly outperforms existing local models and efficiently achieves competitive results with global models.

* Accepted to ACL 2020  
View paper onarxiv icon

Share this with someone who'll enjoy it: