Multitask Easy-First Dependency Parsing: Exploiting Complementarities of Different Dependency Representations

Abstract

We present a parsing model for projective dependency trees which takes advantage of the existence of complementary dependency annotations for a language. This is the case for Arabic with the availability of CATiB and UD treebanks. Our system performs syntactic parsing according to both annotation types jointly as a sequence of arc-creating operations following the Easy-First approach, and partially created trees for one annotation type are also available to the other as features for the score function. This method gives error reduction of 9.9% on CATiB and 6.1% on UD compared to a single-task baseline, and ablation tests show that the main contribution of this reduction is given by sharing tree representation between tasks, and not simply sharing BiLSTM layers as is usually performed in NLP multitask systems.

CamelParser2.0: A State-of-the-Art Dependency Parser for Arabic

“…Due to a different experimental setup and data scope explored by Al-Ghamdi et al ( 2023), we cannot directly compare our results on all metrics and datasets; however, we observe that our approach outperforms their reported results on the test set of PADT. Additionally, by comparing our findings to the existing SOTA pipelines, Camel-Parser1.0 and UDPipe 2, as well as the reported results in Kankanampati et al (2020), we observe that CamelParser2.0 sets the new SOTA in Arabic dependency parsing for both gold and predicted tokenization settings.…”

“…Their results are observed using Malt-Parser, a transition-based model, with a featurebased SVM classifier (Nivre et al, 2006), which differs from the recent neural SOTA models that learn features from the training data implicitly. Kankanampati et al (2020) leverage the Easy-First LSTM-based architecture proposed by Kiperwasser and Goldberg (2016), but experiment with sharing tree representations and BiLSTM layers between CATiB and UD formalisms to achieve significant error reduction on both.…”

“…We do not train on individual CamelTB genres because of the limited amount of data we have; but we report results for them. Similar to Kankanampati et al (2020), we exclude all non-projective trees in the training, but not in the dev and test.…”

Annotations Matter: Leveraging Multi-task Learning to Parse UD and SUD

“…They find that in all settings, performance on the target constituency treebank improves, with the highest gain coming from using all five as an auxiliary treebank. Kankanampati et al (2020) use the Multidimensional Easy First approach introduced by Constant et al (2016) to parse the Arabic CATiB (Habash and Roth, 2009) and its converted UD representation in a multi-task setup. They note that both treebanks showed error reduction, but that improvements were due to partial dependencies, and not primarily driven through lexical sharing.…”

Proceedings of the 17th International Conference on Parsing Technologies and the IWPT 2021 Shared Task on Parsing into Enhanced Universal Dependencies (IWPT 2021)

“…More explicit MTL settings with treebanks representing different individual tasks have proven successful in array of settings across languages and architectures (Guo et al, 2016;Johansson and Adesam, 2020;Kankanampati et al, 2020).…”