Multiword Expressions: A Pain in the Neck for NLP

What it is about

Not read: this card is written from what is reliably known about the paper, and every idea is marked to check against the text. The paper argues that multiword expressions are frequent and that neither treating them as single "words with spaces" nor treating them as fully compositional works. It names two problems: idiomaticity (the meaning or syntax is not predictable from the parts) and overgeneration (a naive compositional grammar licenses variants nobody says). It proposes a classification. Lexicalised phrases split into fixed expressions (no variation at all), semi-fixed expressions (non-decomposable idioms, compound nominals, proper names: inflection but little else) and syntactically flexible expressions (verb-particle constructions, decomposable idioms whose parts carry meaning and can be passivised or modified, light-verb constructions). Institutionalised phrases are compositional but conventional, such as fixed collocations. Each class gets a different treatment in an HPSG grammar.

Ideas

• I1. The class as an enum: every entry of the pragmatics expressions base and the en dictionary's multiword entries gets one of the paper's classes, and the class decides what the parser allows (words-with-spaces only for fixed expressions; inflection for semi-fixed; full syntax with a lexical link for flexible ones). (to check: the exact class names and boundaries)
• I2. Decomposable versus non-decomposable idioms as a test: an idiom whose parts carry meaning allows passive and internal modification, one that does not keeps its literal reading in those frames. Write minimal pairs into 19. The lexicon: vocabulary, multiword expressions and register and test that prag reads the idiom only where the class allows it. Negative control: the literal sentence must not get the idiomatic reading. (to check)
• I3. Light-verb constructions in the frames reader: a light verb plus an eventive noun yields the noun's frame with the verb's subject as agent, so the world records the event, not the light verb. Measure on teacher-checked sentences from the tale lessons. (to check: the paper's treatment)
• I4. Institutionalised phrases are a frequency fact, not a grammar fact: keep them as a counted list with a score and do not give them special syntax. (to check)

For Mova

The chapter is 19. The lexicon: vocabulary, multiword expressions and register; the code is en/src/dict.rs, en/src/parse.rs, and the expression reader and frames in prag (prag/src/frames.rs). The card holds only our summary.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.1007/3-540-45715-1_1

@article{sag2002multiword,
  title = {{Multiword Expressions: A Pain in the Neck for NLP}},
  author = {Sag, Ivan A. and Baldwin, Timothy and Bond, Francis and Copestake, Ann and Flickinger, Dan},
  journal = {CICLing 2002, LNCS 2276 (Springer), 1–15},
  year = {2002},
  doi = {10.1007/3-540-45715-1\_1},
  url = {https://doi.org/10.1007/3-540-45715-1_1}
}

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/idioms-2002-sag-multiword-expressions-pain-neck-nlp/