📄 Symmetric Pattern Matching Analysis for English Coordinate Structures

Summary

The paper addresses a long-standing difficulty for rule-based machine translation: deciding which constituents an English coordinator ("and", "or", "but", commas and similar words) actually joins. Grammar-based approaches tend to produce an explosion of candidate parses, and the problem worsens with part-of-speech ambiguity, unknown words and ellipsis. The authors exploit the tendency of conjuncts to look alike. Each word is described by three feature layers: phrase role, a fine-grained word class, and morphological shape. A "balance matching" procedure then searches for the left and right spans around the conjunction that share the most features. The resulting scope is passed top-down to the parser. Patterns and weights were derived from 10,000 conjunctive sentences in technical manuals. When tested on 15,000 further sentences inside the PIVOT English-Japanese MT system, the module gave correct scope for roughly 75% of them. It nearly eliminated backtracking and helped resolve ambiguous and unknown words. This matters because parallelism proves to be a cheap and robust cue that complements hand-written grammar rules.

Key points

• Each word is represented by three feature sets. The phrase feature takes values such as Predicative, Nominal, Nominal-Premodifier, Nominal-Postmodifier and Predicate-modifier. The word feature has about 120 subclasses of parts of speech, such as NounInstrument or NounHuman. The morphological feature pairs a word with its character type, such as alphabet_hyphen_arabic-numbers.
• Balance matching uses three operations: feature intersection, matching of word-feature sets, and balancing of the two lists. Its goal is to pick the start of the left conjunct and the end of the right conjunct that give maximum shared features.
• Analysis of 10,000 conjunctive sentences from technical manuals yielded about 300 coordinate patterns. The feature weights were set manually as binary values (1.0 or 0.0).
• The evaluation used 15,000 held-out sentences in PIVOT, which has a 100,000-word dictionary. Results:
  • Scope was correct for about 75% of sentences.
  • Backtracking was largely removed.
  • About one sixth of the sentences contained unknown words.
  • Word disambiguation accuracy roughly doubled.
• The authors name three planned improvements: local filtering rules applied before matching, learned real-valued weights, and semantic similarity measures based on a taxonomy.

Interesting ideas

• I1. Conjunct scope is treated as an alignment problem: the system picks the left and right spans whose per-word feature sets overlap most, which turns a parse-explosion problem into one bounded search.
• I2. Ambiguous word classes are kept as sets and intersected with the counterpart word in the other conjunct, so parallelism itself removes noun/verb readings (as with "ditching hatches and escape hatches").
• I3. Character-shape classes (case pattern, letter-hyphen-digit forms such as "uranium-235") act as a fallback feature, so unknown words can still be matched against their counterparts.
• I4. Gapped material is recovered as the set difference between a fuller left conjunct and a shorter right conjunct, which restores the elided verb phrase or head noun.
• I5. Coordination scope is resolved in a cheap pre-pass and handed down as constraints, so the main analyzer avoids costly backtracking.

For Mova

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

• I1 (tree): A symmetry score over candidate conjunct spans could rerank conj attachments from the perceptron parser or gparse, and the aligned features would also serve as a readable justification.
• I2 (tokens): The tagger could keep k-best tag sets inside coordinations and narrow them by intersecting with the aligned twin, logging which counterpart forced each choice.
• I3 (tokens): Shape features for out-of-vocabulary tokens such as codes, model numbers and abbreviations could feed both tagging and coordination decisions when no lexicon entry exists.
• I4 (frames): Gapping constructions could be expanded into complete who-did-what frames by copying the missing predicate and arguments from the aligned conjunct, with the alignment recorded as the explanation.
• I5 (tree): A symmetry-based scope hint could constrain or seed the dependency parser, or serve as one more witness when building silver coordination data.