汉语被动结构解析及其在CAMR中的应用研究

Summary

The paper addresses the automatic analysis of the internal makeup of Chinese passive sentences. Earlier work could only flag whether a sentence was passive. The authors treat the problem as a specialised form of semantic role labeling. Each passive structure is a tuple of verb (V), patient-like argument (A1), optional agent (A0) and optional marker (M), which can be a preposition such as 被 or a "suffering" verb such as 遭到.

They sampled sentences from an existing passive-sentence corpus and labelled them at the word level. The scheme uses BIO tags plus an index that links each component to its verb. Their model, PS-CRF, has three parts:

• BERT-wwm-ext embeddings
• biaffine attention scoring over a tree in which each verb heads its components
• a TreeCRF output layer with Viterbi decoding

PS-CRF reaches an F1 of 97.31%. The parser's output was also used to post-process CAMR graphs produced by SPRING on primary-school Chinese texts, which raised both Smatch and Align-Smatch. This shows that targeted structural parsing can improve broader semantic graph parsing.

Key points

• Dataset: 3,839 passive sentences were annotated with 4,814 passive structures, about 1.25 per sentence.
  • 2,711 are marked: 2,465 with a prepositional marker, 190 with a suffering-verb marker, and 56 attributive cases.
  • 2,103 are unmarked, including a small number of attributive cases.
  • The data was split 6:2:2 into train, dev and test.
• Results: PS-CRF scores P 98.46 / R 96.18 / F1 97.31. Baselines:
  • plain CRF with GloVe: F1 77.59
  • Biaffine+TreeCRF: F1 80.47
  • BERT+Biaffine+TreeCRF: F1 96.44
• Why the tree helps: Modelling each passive structure as its own dependency tree lets the model handle sentences with several, possibly nested, passive structures. A flat CRF cannot do this.
• Remaining errors: Most come from attributive (定中) passive structures, which are under 2% of the data. Others come from words with several senses that get mistaken for passive verbs.
• CAMR post-processing with SPRING:
  • Without alignment: Align-Smatch rose from 45.2 to 47.6 and Smatch from 69.8 to 70.74.
  • With manually added concept alignment: Align-Smatch rose from 62.3 to 65.8.
  • The Smatch gains are smaller because markers are encoded inside relation labels rather than as separate nodes.