Acquiring and Generalizing Causal Inference Rules from Deverbal Noun Constructions

Shohei Tanaka, Naoaki Okazaki, Mitsuru IshizukaCOLING 2012 · OpenReview

Abstract

Researchers have developed a novel way to teach computers "if-then" causal logic by leveraging deverbal nouns (nouns derived from verbs, like acquisition from acquire). By combining these linguistic clues with large-scale information extraction and machine learning, the system induces complex causal rules that include necessary context—moving beyond simple "A causes B" to more nuanced "A causes B if C is true."

Introduction

For an AI to perform tasks like Question Answering or Discourse Analysis, it requires a library of commonsense causal knowledge. While humans intuitively understand that "Company A acquiring Company B" might lead to "Company A competing in a new market," teaching this to a computer is difficult. Previous research focused on simple, one-to-one mappings (e.g., hunger $\Rightarrow$ headache). However, real-world causality is often conditional. Generalizing "Google acquires Android" into "X acquires Y $\Rightarrow$ X competes in Z" is too vague. The challenge lies in finding "hidden" constraints—specific conditions like "Y is specialized in Z"—that make a causal rule valid.

Methods

The authors propose a multi-stage pipeline to extract and refine these rules:

  1. The Deverbal Clue: The system uses FrameNet to identify pairs of verbs and their noun forms. It looks for sentence pairs where a verb is referred back to by its noun form (e.g., "UnitedHealth buys Pacificare. The acquisition also gives..."). This structure serves as a high-precision indicator of a cause-effect relationship.
  2. Pattern Simplification: Using dependency parsing, the system identifies core arguments (Subject, Object, Location) to turn specific sentences into generalized templates.
  3. Constraint Discovery: To solve the "too generic" problem, the system queries ReVerb against the ClueWeb09 corpus to find relations that connect variables.
  4. The Ranker: A binary classifier using a Linear Kernel SVM selects the most plausible constraints based on features like word overlap and document similarity.
The process of identifying head and body patterns and selecting candidates for constraints.

Key Insights

  • Deverbal nouns as anchors: When a writer uses a deverbal noun to refer to a previous verb, they are typically describing consequences. This is a higher-precision starting point than simple proximity-based extraction.
  • The "Hidden Predicate" Problem: Useful causal rules often depend on information not explicit in the text (e.g., knowing a company's specialization). By using ReVerb to find these "hidden" predicates, the system mimics human background knowledge.
  • Selectivity over Connectivity: While simple baselines connecting any two events have high recall, they suffer from abysmal precision. The deverbal noun approach acts as a powerful filter for high-confidence causal links.

Results

The proposed method achieved a Precision of 0.5357 in locating causal relations, significantly outperforming the baseline method's 0.0445. This confirms that deverbal nouns are excellent clues for causality.

Comparison of precision, recall, and F1 scores between the baseline and proposed methods.

In terms of rule accuracy, the "Proposed method + SVM" achieved an accuracy of 0.375 for pure causality. While low in absolute terms, it more than doubled the performance of the "ReVerb+ReVerb" baseline (0.1667), proving the necessity of the constraint-selection step.

Conclusion

This research demonstrates that deverbal noun constructions are effective anchors for extracting nuanced causal rules. Future work will focus on paraphrasing raw text fragments into clean predicate-style logic and addressing data sparsity to capture causal relations expressed through different rhetorical devices.

Summary updated

BibTeX
@inproceedings{tanaka2012acquiring,
  author = {Shohei Tanaka and Naoaki Okazaki and Mitsuru Ishizuka},
  title = {{Acquiring and Generalizing Causal Inference Rules from Deverbal Noun Constructions}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2012},
  url = {https://openreview.net/forum?id=FCHP8oHz3oJQ},
}

Loading related content...

Figure