📄 A Noun Phrase Parser of English

Summary

Voutilainen presents a rule-based noun phrase parser for English, built in the Constraint Grammar framework (Karlsson 1990). It is meant to give shallow but highly reliable analyses for tasks such as index term generation, collocation extraction and preprocessing for deeper parsers. The front end reuses the ENGCG pipeline: preprocessing, ENGTWOL two-level morphology, heuristics for unknown words, and constraint-based morphological disambiguation. On top of this sits a new, deliberately lean syntactic scheme of seven function tags. These tags contrast noun phrases with everything else, and heads with modifiers. Parsing is reductionistic. Each word starts with every possible tag, and about 120 hand-written negative constraints remove the readings that are illegitimate in context. On 3,600 words of unseen, partly hard text, 93.5% of words ended up syntactically unambiguous and 99.15% kept the appropriate reading. The author argues that lexical knowledge can live inside rule-based grammars rather than only in statistical preferences, and that this yields a better recall/precision trade-off.

Key points

• The seven tags are @V (verbs and infinitive to), @NH (nominal head), @>N (determiner or premodifier), @N< (unambiguous postmodifying PP), @AH (adjectival heads, adverbials and attachment-ambiguous PPs), @CC and @CS.
• ENGTWOL uses a 56,000-entry lexicon based on Quirk et al. (1985) and gives roughly two readings per word.
  • The morphological disambiguator applies about 1,300 constraints (1,100 grammar-based and 200 heuristic).
  • After it runs, 96–98% of tokens are unambiguous and at least 99.6% keep the correct reading.
• The syntactic grammar has about 120 constraints, around 50 of them heuristic. They express four general statements:
  • part of speech fixes the order of modifiers;
  • only likes coordinate;
  • every modifier has a head;
  • an auxiliary precedes a main verb.
• In the test, 31 correct readings were lost: 18 by syntactic constraints, 11 by disambiguation constraints and 2 by the lexicon. Multi-word adjectives such as language specific were the worst single source of syntactic errors.
• The whole system runs at about 400 words/s on a SPARCstation 10/30, and disambiguation plus syntax alone runs at over 1,000 words/s. The parser already powers NPtool.

Interesting ideas

• I1. Each word first receives every reading a lexicon can propose, and contextual rules only eliminate readings, so the parse is whatever survives and every removal is traceable to a rule.
• I2. The tag set deliberately leaves structurally undecidable distinctions implicit, such as whether fat modifies butcher's or wife, rather than forcing a guess.
• I3. PP attachment is committed only in reliably decidable configurations, such as of-phrases and preverbal NP-PP sequences. All other PPs get a neutral tag that remains easy to locate later.
• I4. Broad structural rules often fail on a small, predictable set of lexical exceptions, such as specific and based after nouns. The author suggests mining such exception lists from large tagged corpora and adding them as lexical context conditions.
• I5. Evaluation reports two figures: how much ambiguity remains and how often the correct reading survives. Error analysis also traces domino effects, where one wrong discard triggers others.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): the tagger could keep the full candidate set from the en lexicon and log which constraint discarded each reading, giving per-token explanations.
• I2 (tree): the parser could mark certain premodifier attachments as underspecified, so frames and world steps do not treat an arbitrary bracketing as fact.
• I3 (frames): prag could treat PPs flagged as attachment-uncertain as open slots and resolve them with VerbNet role expectations instead of trusting the tree.
• I4 (learn): self-induced repair rules could pair each general constraint with an automatically mined exception lexicon taken from silver-data disagreements.
• I5 (learn): parser evaluations could report retention of the gold analysis alongside residual ambiguity, and group cascaded errors under their root cause.