Building a Hebrew Semantic Role Labeling Lexical Resource from Parallel Movie Subtitles

What it is about

The authors build the first Hebrew semantic role resource by projecting English SRL, produced by SEMAFOR, onto aligned Hebrew sentences from OpenSubtitles 2016. The pipeline filters the 23.7M English-Hebrew pairs by language identification (22.4M left), parses Hebrew with YAP after segmenting clitics, computes word alignments with fast_align, and projects spans along the dependency tree. Pairs with one-to-many alignments are dropped; 6.8M pairs survive the projection filter. A linear classifier, trained on 124 sentences that were labelled by hand as good or as one of five error types, keeps pairs scored above 80% good. The FrameNet core roles are mapped to PropBank arguments, giving 423K sentences (240K train, 30K dev, 30K test). A multilingual BERT tagger reaches F1 0.64 with segmented Hebrew and 0.63 unsegmented (Table 2), so segmentation does not hurt.

Ideas

• I1. Filtered projection for Hebrew: keep a projected predicate-argument annotation only when the aligned Hebrew head is a verb (verb filter) and is a translation of the English predicate (translation filter); count the drop rate per filter before any Hebrew annotation is used.
• I2. A quality gate for projected pairs: a small classifier over alignment counts (one-to-one, one-to-many, sentence-length ratio, parse depth), trained on a few hundred hand labels with the six classes (sentence alignment error, poor translation, word alignment error, poor syntactic parsing, poor frame parsing, good); report its precision on held-out labels.
• I3. Report type counts before and after clitic segmentation (the paper gives 2.4M word types before and 0.89M after, with 782K types below ten occurrences) as a standard number for the tools/hbo segmenter, and run the same segmented/unsegmented comparison on any new Hebrew tagger.

For Mova

I1 and I2 belong to the building of treebanks and frames: projection from en to Hebrew can feed candidate annotations to treebanks/he/ (htb, iahltwiki), never as gold; I2's labels go into the quality tooling next to prag. I3 is a measurement for tools/hbo's segmenter. The FrameNet-to-PropBank mapping by core-role order is a note for prag frame roles (keep the core roles only). Paper licence is NC: this card is a summary, and no text or data from the paper goes into a release.

Topics: Semantic role labeling

Cite

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

@inproceedings{eyal2020building,
  title = {{Building a Hebrew Semantic Role Labeling Lexical Resource from Parallel Movie Subtitles}},
  author = {Eyal, Ben and Elhadad, Michael},
  booktitle = {Proceedings of the Twelfth Language Resources and Evaluation Conference},
  year = {2020},
  url = {https://aclanthology.org/2020.lrec-1.727/}
}

References

1 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Akbik, A., Chiticariu, L., Danilevsky, M., Li, Y., Vaithyanathan, S., and Zhu, H. (2015). Generating high quality proposition Banks for multilingual semantic role find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/lingbulk-2020-eyal-building-hebrew-semantic-role-labeling/