The Role of Modifier and Head Properties in Predicting the Compositionality of English and German Noun-Noun Compounds: A Vector-Space Perspective

What it is about

The paper asks which properties of a noun-noun compound's modifier and head make its degree of compositionality easier to predict from a vector space. It uses co-occurrence vectors (sentence-internal window of 20 words each side, local mutual information) from the COW corpora and the cosine between compound and constituent. The new German gold standard, GhoSt-NN, has 180 compounds (balanced) and 868 compounds (extended), rated for compositionality on a 1–6 scale. For English, the compound-head prediction reaches Spearman ρ = 0.60 on the Reddy set (90 compounds), while the both-constituent score on the 1,443-item Ó Séaghdha set is only 0.35. Compounds with low-frequency, low-productivity and low-ambiguity heads are predicted better; the ambiguity effect is significant in the extended GhoSt-NN set (p < 0.01). Relation types also differ, with BE compounds predicted best and ACTOR compounds worst.

Ideas

• I1. en/src/dict.rs: the head's number of senses, from the lexicon, as a gate on the compositional reading. Heads with more than two senses get a "lexicalised" flag. Measure on English compounds with the standard ratings, if a licensed set can be used; control with a random flag.
• I2. world/src/dictionary.rs: the same ambiguity feature on pupil compounds ("birthday cake", "sunflower"). Check that the compound's sense is chosen from the correct head sense in the school questions.
• I3. Relation classes as rules in prag: the relation (for example "made of", "for", "about") from prepositional and verbal cues, with per-class accuracy reported. Keep the relation enum deterministic.
• I4. A count-based replacement for the vector cosine: pointwise mutual information over corpus counts in BTreeMap tables, run twice to confirm identical output.

For Mova

The English chapter is languages/en/18-compounding-conversion-minor.md, and the lexicon is en/src/dict.rs. The paper's results are correlations, not a rule, so Mova would need its own gold data. The German data are not for Mova now. The clearest English point is the head-ambiguity effect, which can be measured with the existing sense table.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/s16-2020

@inproceedings{walde2016role,
  title = {{The Role of Modifier and Head Properties in Predicting the Compositionality of English and German Noun-Noun Compounds: A Vector-Space Perspective}},
  author = {Walde, Sabine Schulte im and Hätty, Anna and Bott, Stefan},
  booktitle = {Proceedings of the Fifth Joint Conference on Lexical and Computational Semantics},
  year = {2016},
  doi = {10.18653/v1/s16-2020},
  url = {https://doi.org/10.18653/v1/s16-2020}
}

References

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

1. Marco Baroni, Silvia Bernardini, Adriano Ferraresi, and Eros Zanchetta. 2009. The WaCky Wide Web: A Collection of Very Large Linguistically Processed Web-Crawled Corpora. Language Resources and Evaluation, 43(3):209–226. find
2. Marco Baroni, Raffaella Bernardi, and Roberto Zamparelli. 2014. Frege in Space: A Program for Compositional Distributional Semantics. Linguistic Issues in Language Technologies, 9(6):5–110. find