Elamur Research
☰
home / papers / 2019 / Joint Transition-Based Models for Morpho-Syntactic Parsing: Parsing Strategies for MRLs and a Case Study from Modern Hebrew

Joint Transition-Based Models for Morpho-Syntactic Parsing: Parsing Strategies for MRLs and a Case Study from Modern Hebrew

What it is about

In Hebrew, morphological analysis and disambiguation (MA&D) normally comes before parsing, yet the syntax would help choose the morphology, and the reverse. The authors merge a morphological transition system (More and Tsarfaty 2016) with two dependency systems (Arc-Standard and Arc-ZEager) into one transition-based parser, with one score, joint training and beam search (beam 64). They compare three strategies: a pipeline; MDFirst, where the parser sees all beam candidates of the morphology; and ArcGreedy k, where syntax moves ahead once k morphemes are in its buffer. An ENDTOKEN transition handles lattice paths of different lengths. On the Modern Hebrew SPMRL data, with gold paths infused into the lattices, morphological F1 rises from about 93.3 (standalone) to about 94.4–94.7 under the joint strategies, while dependency F1 drops by up to 0.5 point. Arc-ZEager is slightly better than Arc-Standard. The joint system beats earlier Hebrew parsers end to end.

Ideas

For Mova

I1 goes into tools/hbo (segmenter to gparse) and languages/hbo/13-mova-tools.md. I2 and I3 go into the tools/hbo evaluation and languages/he/11-ud-analysis.md. The paper is CC BY 4.0, so a short phrase may be quoted with its page; no sentences are copied.

Library summary

The Research library's own summary of the paper, in our words — not the paper's text.

Written for the Research library (claude-opus-5-5) from the paper text. Original — aclanthology.org/Q19-1003.pdf

Summary

In morphologically rich languages (MRLs) such as Modern Hebrew, one space-delimited token can contain several syntactic units, and unvocalized script makes segmentation highly ambiguous. Standard pipelines run morphological analysis and disambiguation (MA&D) before parsing, so segmentation errors propagate, and local context alone often cannot resolve them. The authors test the long-standing hypothesis that morphology and syntax should be disambiguated together. They build a language-agnostic transition-based parser that merges a morphological transition system, which picks a path through a lattice of candidate analyses, with a dependency transition system. The combined system uses one objective, joint training with a structured perceptron, and joint beam-search decoding. On the Hebrew SPMRL data, joint prediction gives better MA&D than the standalone morphological component and sets a new state of the art for end-to-end Hebrew dependency parsing. It surpasses both pipeline parsers and an earlier joint graph-based system. The work shows that syntax helps morphology in realistic settings, and the authors release the parser as open source.

Key points

  • The morphological part extends the More and Tsarfaty (2016) lattice-based MD system, previously the best for Hebrew MA&D. It adds an ENDTOKEN transition to handle lattice paths of varying length during beam search; in the paper's example, path lengths range from 4 to 7 arcs.
  • Learning uses the early-update averaged perceptron (Collins and Roark, 2004) within the Zhang and Clark (2011) structure-prediction framework, with beam-search decoding.
  • The joint parser clearly beats the dual-decomposition graph-based joint model of Seeker and Çetinoğlu (2015). It also substantially outperforms Goldberg and Elhadad (2010), long treated as the default Hebrew dependency parser.
  • A qualitative error analysis attributes the remaining errors to several sources:
    • clause and PP attachment
    • segmentation or tagging mistakes
    • Hebrew-specific phenomena, such as unmarked indefinite objects and construct-state nouns
    • inconsistent or incorrect labels in the treebank itself
  • The paper is published in TACL vol. 7 (2019) under CC-BY 4.0. Future plans include neural morphological and syntactic models that can interact with each other.

Topics: Dependency parsing

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.1162/tacl_a_00253

@article{more2019joint,
  title = {{Joint Transition-Based Models for Morpho-Syntactic Parsing: Parsing Strategies for MRLs and a Case Study from Modern Hebrew}},
  author = {More, Amir and Seker, Amit and Basmova, Victoria and Tsarfaty, Reut},
  journal = {Transactions of the Association for Computational Linguistics},
  year = {2019},
  doi = {10.1162/tacl\_a\_00253},
  url = {https://doi.org/10.1162/tacl_a_00253}
}

References

4 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Steven P. Abney and Mark Johnson. 1991. Memory requirements and local ambiguities of parsing strategies. Journal of Psycholinguistic Research, 20(3):233–250. find
  2. Meni Adler. 2007. Hebrew Morphological Disambiguation: An Unsupervised Stochastic Word-Based Approach. Ph.D. thesis, Ben-Gurion University of the Negev, Israel. find
  3. Meni Adler and Michael Elhadad. 2006. An unsupervised morpheme-based HMM for Hebrew morphological disambiguation. In Proceedings of ACL. Association for Computer Linguistics. find our card
  4. Daniel Andor, Chris Alberti, David Weiss, Aliaksei Severyn, Alessandro Presta, Kuzman Ganchev, Slav Petrov, and Michael Collins. 2016. Globally normalized transition-based neural networks. In Proceedings of the ACL. Association for Computational Linguistics. find our card