Learning to Identify Reduced Passive Verb Phrases with a Shallow Parser

Sean Igo, Ellen RiloffAAAI 2008

Introduction: The Challenge of Reduced Passives

In Natural Language Processing (NLP), correctly assigning thematic roles—identifying who did what to whom—is essential. While ordinary passives (e.g., "YouTube was purchased by Google") are relatively easy for modern systems to identify, English frequently uses reduced passives. In these constructions, the auxiliary verb "to be" is omitted (e.g., "the man arrested"), often causing parsers to incorrectly label the subject as the agent of the action rather than the theme.

While full syntactic parsers can identify these structures, they are often computationally expensive and struggle with domain portability. As shown in the performance data, even top-tier parsers see a significant drop in accuracy when moving from standard financial news (WSJ) to more specialized domains like terrorism reports (MUC4) or medical data (PRO).

Performance of standard full parsers (CH, CO, MP) on ordinary vs. reduced passives across three distinct domains.

Methodology

The authors propose a machine-learning framework to bridge the gap between fast shallow parsers and the deep structural analysis required for reduced passives. Their approach involves:

  1. Candidate Filtering: Identifying verbs that are potential reduced passives based on tense and the absence of auxiliary verbs.
  2. Feature Engineering: A suite of 28 features was developed, including lexical/syntactic cues, Part of Speech (POS) tags, and semantic classes provided by the Sundance shallow parser.
  3. Novel Signal Features: They introduced empirical transitivity checks (how often a verb requires an object) and thematic role semantics (how likely a noun is to be a "theme" vs. an "agent" for a specific verb).
  4. Classifiers: The study compared Decision Trees (J48), Linear SVMs (LSVM), and Polynomial Kernel SVMs (PSVM).

Key Results

The Polynomial SVM (PSVM) consistently outperformed other methods across all tested domains. While simple rule-based heuristics (like checking for a "by" phrase) achieved high precision, they suffered from abysmal recall—often as low as 9%. In contrast, the PSVM maintained a robust balance, achieving 80-84% precision and roughly 60-65% recall.

Comparison of the proposed machine learning models (Dtree, LSVM, PSVM) against baseline heuristics across multiple datasets.

Beyond accuracy, the shallow parser approach offered a massive efficiency gain. The system processed 1,718 sentences in just 20 seconds, whereas a standard full parser required 48 minutes for the same task. This makes the ML-augmented shallow parser a viable solution for real-time or large-scale text processing where full syntactic parsing is impractical.

Conclusion

This research demonstrates that specialized machine-learning layers can empower fast, robust shallow parsers to handle complex linguistic phenomena like reduced passives. While a recall gap remains, the system provides a significant improvement over standard heuristics and maintains high precision across diverse text domains. Future work will focus on refining semantic features and morphological analysis to further improve the identification of these tricky verb phrases.

Summary updated

BibTeX
@inproceedings{igo2008learning,
  author = {Sean Igo and Ellen Riloff},
  title = {{Learning to Identify Reduced Passive Verb Phrases with a Shallow Parser}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
}
Figure