汉语增强依存句法自动转换研究

Summary

Basic dependency trees describe surface syntax, but they often hide relations between content words. Typical gaps include arguments shared by coordinated items, roles altered by passive or 把 constructions, and elided subjects. Enhanced representations already exist for English (enhanced UD, enhanced++, pyBART/BART), but Chinese had no comparable scheme. The authors build a Chinese enhanced dependency guideline on top of Stanford basic dependencies. It adds arcs and enriches labels in four areas:

• lexicalising prepositions and conjunctions;
• propagating conjuncts and appositives;
• normalising alternations toward active-voice relations;
• adding arguments in pivotal, serial-verb and elliptical sentences.

Unresolvable cases are flagged with the UNC and ALT markers borrowed from pyBART. A rule-based Python converter turns basic trees into enhanced graphs. A web demo shows the basic and enhanced analyses side by side. Two case studies, collocation extraction and information extraction, show that the enhanced graphs recover relations the basic trees miss. The paper reports no quantitative evaluation such as LAS or accuracy. Its value lies in the openly released resource and in its potential use for corpus-based teaching and IE in Chinese.

Key points

• Preposition and conjunction labels: prepositions are folded into labels (e.g., nmod:prep_向), with circumfix prepositions written as both parts plus an ellipsis (nmod:prep_在……上). Coordinating conjunctions are folded the same way (conj_和, conj_或者).
• Alternation normalisation:
  • Passive subjects, 把 objects and heads of passive or object-less verbal modifiers receive an extra dobj arc.
  • Adjectival modifiers and some verb-phrase modifiers point to their head noun via nsubj.
  • Pivotal (兼语) and serial-verb sentences get added nsubj arcs.
  • Elided subjects and objects are restored where possible.
• Uncertainty handling: arcs whose correctness the user must judge are marked UNC=TRUE, and arcs with several possible antecedents are marked ALT=X. Both ideas come from Tiktinsky et al.'s pyBART work.
• Released resources: the converter relies on POS tags, arc span and direction, and dependency labels. Code is on GitHub (blcuicall/chinese-enhanced-dependencies), and the online demo at parser.blcu-icall.org gives examples for 17 enhancement rules.
• Applications and future work: collocations are extracted via nsubj, dobj, advmod:dvp and compound:nn, and enhanced graphs recover collocations of coordinated items that basic trees miss. Planned future work includes judgment and inverted sentences, plus transitivity-aware handling of elided objects.