📄 Linear Neural Parsing and Hybrid Enhancement for Enhanced Universal Dependencies

Summary

The paper describes the University of Pisa system for the IWPT 2020 Shared Task. The task was to produce enhanced UD graphs from raw text for 17 languages. The authors focused on two stages: syntactic parsing and dependency enhancement. Sentence splitting, tokenization and tags were taken from UDPipe. The base parser chosen was the transition-based UUParser, which runs in linear time. It was extended with BERT/AlBERT embeddings and with features from syntactic probes that estimate the distance between words in the tree. The official submission is an ensemble of UDify, UUParser and UDPipe, combined with a linear tree-combination algorithm. A later unofficial five-parser version also included the biaffine parsers Stanza and Zysite. On average, it trailed the winner by only 1.88 UAS. For enhanced dependencies, the authors combined rules and learning. They took a modular, parameterized and language-independent graph conversion script. For difficult cases, they complemented it with a multi-task classifier. The work shows that fast linear-time parsers can compete with quadratic graph-based ones. It also shows that different treebanks follow the enhanced UD guidelines to differing degrees.

Key points

• Average speed across all 17 test sets on GPU: Zysite — 84.82 sentences/s, UUParser — 16.62, Stanza — 7.77. Zysite was also the most accurate (UAS 86.67 vs. 83.93 for UUParser).
• Hints from the syntactic probe slightly improved UUParser on the English dev set: 92.21 UAS / 90.31 LAS instead of 91.32 / 89.33 without them.
• The jointly trained classifier makes three binary decisions. It identifies the external subject in xcomp constructions and decides whether to propagate the head or the dependents to conjuncts. For each language, the choice of whether to use the classifier (ml) and the label mapping (patch), and which enhancement types to exclude, was tuned separately on the dev set.
• The weakest results are on the Baltic languages, Finnish and Arabic. Even the ensemble trails the best submission on Arabic by 8.2 UAS. From EULAS to ELAS the system loses 6.26 points on average because it predicts labels carrying case or preposition information poorly.
• The official submission is named UNIPI-003, the unofficial one UNIPI-comb5. The code of the extended UUParser is published on GitHub: attardi/uuparser.