PP Attachment: Where do We Stand?

What it is about

Prepositional phrase (PP) attachment is one of the most frequent parser errors: in the authors' German parser about 18% of labelled attachment errors are PP attachments. The paper treats attachment as a choice among all candidate heads of a preposition (3.15 on average), scored by a feed-forward network. The inputs are word and tag embeddings, topological-field features from the clause structure, and auxiliary bilexical preference distributions (pointwise mutual information) estimated from a parse of 28.8 million sentences of newspaper text. On a German data set of about 44,000 instances, the accuracy rises from 56.9% for logistic regression to 86.5% with all features and 86.7% with auxiliary counts from unambiguous attachments only. The error analysis puts the ceiling near 92.6%: most remaining errors are ambiguities that the gold standard marks as one of several correct attachments, or that need discourse context.

Ideas

• I1. en/src/parse.rs: treat PP attachment as a candidate choice for obl, nmod and advcl arcs with a PP head, and measure the attachment accuracy on EWT test and the share of labelled errors it accounts for (the paper finds 18% in German).
• I2. Auxiliary bilexical counts (PMI of preposition-object-head triples) from the school silver corpus (annot, 1.17M words) and from the en parses of a larger raw collection; metric: accuracy on EWT PP attachments with and without the counts. The data is English, not the German set of the paper.
• I3. Ceiling check first: count the EWT PP attachments that have two or more defensible heads, as the paper's ceiling argument does, and report the accuracy ceiling per treebank before tuning any model.

For Mova

The chapter is 22. Phrases: noun, adjective, adverb and prepositional phrases, with 23-complementation.md for the verb-or-noun choice. The code is en/src/parse.rs (the candidate set) and en/src/ctx.rs (document context). The German numbers are not Mova's numbers; any English figure must come from EWT test.

Cite

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

@inproceedings{kok2017pp,
  title = {{PP Attachment: Where do We Stand?}},
  author = {Kok, Daniël de and Ma, Jianqiang and Dima, Corina and Hinrichs, Erhard},
  booktitle = {Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers},
  year = {2017},
  url = {https://aclanthology.org/E17-2050/}
}

References

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

1. Table 2: Error analysis of a random sample of 100 PPs that are incorrectly attached by the best model.* Yonatan Belinkov, Tao Lei, Regina Barzilay, and Amir Globerson. 2014. Exploring compositional architectures and word vector representations for prepositional phrase attachment. Transactions of the Association for Computational Linguistics, 2:561–572. find our card

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/onto-2017-kok-pp-attachment-where-do-we/