📄 A Syntactic Analysis Method of Long Japanese Sentences Based on the Detection of Conjunctive Structures

Summary

Kurohashi and Nagao address a classic weakness of Japanese parsers: long sentences, especially those over 50 characters, often break down because coordination (conjunctive noun phrases, coordinated predicate clauses including the renyoh chuushi-ho form, and "incomplete" coordinations) produces heavy scope ambiguity and frequent ellipsis. Earlier approaches only looked for the single word most similar to the word before the conjunction. This paper instead compares whole spans of bunsetsus on both sides of a key bunsetsu (KB). It uses dynamic programming over a matrix of pairwise bunsetsu similarities to find the best-matching pair of conjuncts. Overlapping conjunctive structures are then checked and re-detected if they conflict. Each detected structure is collapsed into one node, and the reduced sentence is parsed with simple head-dependent (kakari-uke) rules. The same dependency pass identifies and restores components omitted inside conjuncts. Scoring weights were tuned by hand on 30 training sentences, and evaluation used 150 Japanese sentences. About 65% of sentences received a fully correct dependency structure, and most of the remaining errors left the main coordination skeleton intact.

Key points

• The authors observe that many existing analyzers produce poor results above about 50 Japanese characters and almost always fail above about 80.
• Coordinations fall into three types: conjunctive noun phrases, conjunctive predicative clauses, and incomplete conjunctive structures. Each type is signalled by a KB, which carries a marker word or a renyoh form.
• The method has three stages:
  • detecting the scope of each conjunctive structure by maximizing span-level similarity with dynamic programming;
  • resolving incorrect overlaps between multiple structures;
  • running dependency analysis inside each conjunct and then on the reduced sentence.
• Parameters were set by hand on 30 sentences, and testing used 150 sentences. Roughly two-thirds of the test sentences were parsed completely correctly.
• The error analysis identifies four main causes:
  • highly unbalanced conjuncts;
  • extra modifiers that appear only in the pre-conjunct;
  • part-of-speech similarity that is too coarse;
  • technical terms that could only be compared by matching characters, with no thesaurus available.

Interesting ideas

• I1. Treat coordination as span alignment. Score whole candidate conjunct pairs using a dynamic-programming path through a pairwise similarity matrix, rather than matching single heads. The same alignment idea could help with coordination or parallel-clause detection in other languages, including Ukrainian.
• I2. Reduce first, then parse. Collapsing detected coordinations into single nodes leaves a short sentence that simple, robust rules can handle. This kind of preprocessing could shorten inputs for any parser that struggles with sentence length.
• I3. Get ellipsis recovery from the same pass. Omitted components inside conjuncts are found as a side effect of the dependency analysis, not by a separate module. That suggests gapping and shared arguments can be handled while building dependency trees.
• I4. Use a balance prior with penalties. The method assumes conjuncts have roughly similar length and penalizes spans that cross strong boundaries. It is a cheap structural bias, and the paper's own error analysis shows where it fails: very unbalanced coordinations.
• I5. Judge the skeleton, not only the full sentence. Evaluating whether coordination scopes and the core structure are right, alongside full-sentence correctness, shows useful partial success that a strict sentence-level score would hide.