Parsing Run Amok: Relation-Driven Control for Text Analysis

Paul S. JacobsAAAI 1992

Abstract

Traditional natural language parsers often struggle with complex, real-world text due to a combinatorial explosion of syntactic structures. This paper introduces relation-driven control, a method implemented in the GE NLToolset that prioritizes identifying key semantic relations over building exhaustive sentence trees. By focusing on linguistic relations—triples of head, role, and filler—the system achieves greater speed and robustness in large-scale data extraction tasks.

Introduction: The Problem of "Clause-Driven" Parsing

In the early 1990s, Natural Language Processing (NLP) transitioned from analyzing simple, isolated sentences to extracting data from massive volumes of real-world text, such as daily news feeds. Traditional syntactic parsers faced a fundamental flaw: they were too "clause-driven." As parsers added more rules to handle linguistic complexity, they encountered a "double-edged sword of coverage," where increased rules led to more misinterpretations.

Specifically, parsers suffered from the "entrapment of attachment"—spending excessive computational effort deciding the exact attachment of modifiers (e.g., prepositional phrases) even when such decisions did not alter the core meaning of the extracted data.

Methodology: Relation-Driven Control

The authors propose a shift toward relation-driven control. Instead of requiring a complete, top-to-bottom syntactic tree, the system focuses on identifying linguistic relations—triples consisting of a head, a role, and a filler.

System architecture of the GE NLToolset featuring relation-driven control.

The system architecture is divided into three main phases:

  1. Pre-processing: The text is segmented into chunks, and "template activators" (typically verbs) are identified to narrow the search space.
  2. Analysis: This phase involves parsing, sense discrimination, and role mapping, all overseen by the relation-driven control mechanism.
  3. Post-processing: Includes confidence scoring, role extension, and default filling.

Semantic Preferences

To validate relations, the system utilizes three types of weighted scores to ensure semantic coherence:

  • Role Preference: Evaluates how well a filler fits a specific role (e.g., "food" as a "patient" for "eat").
  • Rel Preference: Measures how much a head "expects" a certain role.
  • Base Preference: Captures general lexical associations between words.
The R-OBJECT structure illustrating the head-role-filler relation and associated semantic preferences.

This approach allows for fault tolerance; the system can extract useful information even when sentence structures are too complex for a traditional full parse. By focusing on these core relations, the parser avoids getting lost in the syntactic minutiae that do not impact the final data extraction.

Evaluation and Datasets

The system was evaluated using two primary benchmarks:

  • MUC-3 (Message Understanding Conference): A dataset of news reports regarding terrorism in Latin America, used to test information extraction performance.
  • Wall Street Journal (WSJ) Corpus: Used to analyze typical sentence complexity and test the limits of parser coverage.

Key Insights

  • Attachment is Overrated: In data extraction, the exact syntactic attachment of a modifier often doesn't change the resulting data template. If the system knows a "bombing" occurred in a "town," it often doesn't need to decide if the location modifies the "attack" or the "destruction."
  • Pivoting on Verbs: By focusing on domain-specific "pivots" (like verbs related to kidnapping or mergers), the system ignores irrelevant text and focuses linguistic energy where it matters most.
  • "Weak" Method Strength: The paper argues that statistical or domain-driven "weak" methods should control linguistic parsing rather than replace it. Parsing should be a tool used selectively.

Results and Future Work

The relation-driven approach achieved a processing speed of over 1,000 words per minute, a 12x improvement over previous software iterations. In MUC-3 evaluations, the GE system was a top performer, producing significantly less "noise" than competitors and successfully handling sentences over 30 words long—lengths that typically cause traditional parsers to crash.

Future work involves automating the assignment of preference weights, which were largely manual at the time of the study, and integrating these control strategies with more established multi-lingual parsing algorithms, such as the CMU Generalized LR Parser.

Summary updated

BibTeX
@inproceedings{jacobs1992parsing,
  author = {Paul S. Jacobs},
  title = {{Parsing Run Amok: Relation-Driven Control for Text Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1992},
}
Figure