📄 Corpus Based PP Attachment Ambiguity Resolution with a Semantic Dictionary

Summary

The paper addresses prepositional phrase (PP) attachment: deciding whether a PP modifies the verb or the object noun in verb–noun–preposition–noun quadruples. Earlier corpus methods, such as the Collins and Brooks backed-off model, rely on exact word matches and therefore suffer from sparse data. The authors replace string matching with a semantic distance measure over WordNet. This distance combines path length to the nearest common ancestor with concept depth. Because no large sense-tagged parsed corpus was available, they first annotate the training quadruples from the parsed corpus [MA93] with WordNet senses. They do this with an unsupervised, context-similarity word sense disambiguation procedure. A decision tree is then induced over the sense hierarchy to classify attachments. Its leaves also yield a certainty score. The system reaches about 88.1% accuracy, which is roughly equal to human performance when people see only the quadruple. This suggests that lexical-semantic generalisation can substantially offset missing world knowledge in syntactic disambiguation.

Key points

• Reference points:
  • Using each preposition's most frequent attachment gives a 72.7% baseline.
  • Humans reach 88.2% on bare quadruples and 93.2% with full sentence context.
  • The previous best, the Collins and Brooks backed-off model, scored 84.5%.
• Semantic distance: the measure averages the ratios of path length to depth for both concepts in WordNet.
  • Identical concepts score 0.
  • Concepts with no shared ancestor score 1.
• Decision tree: expansion stops once more than 77% of a node's examples share one class. Errors tend to cluster in low-certainty decisions.
• Results: overall accuracy is 88.1% on about 3,100 test quadruples. Per-preposition learning curves are unstable below about 1,000 training examples and then plateau at different levels.
• Future work: the authors expect further gains from wider sentential context, more training data, and more accurate sense tagging.