Investigation of Processing Strategies for the Structural Analysis of Arguments

Robin CohenACL 1981 · DOI

Introduction

In the early 1980s, a foundational effort in computational linguistics sought to move beyond simple sentence parsing toward the structural analysis of entire arguments. This research proposed a framework for computationally reconstructing the logical structure of one-sided arguments by modeling them as trees where claims act as "fathers" to their evidence.

The Problem of Computational Argumentation

When humans listen to an argument, they instinctively build a mental map of which statements support which claims. For a computer, this task is notoriously difficult. Earlier "unsophisticated" models attempted to find these links by testing every new statement against every previous statement, leading to a "combinatorial explosion" of possibilities.

The core challenge addressed is the need for a selective processing strategy: a method to limit the search to only the most likely "logical neighbors" of a new proposition, making the understanding process computationally feasible.

Methodology: Arguments as Trees

The proposed methodology treats an argument as a tree structure. In this model, a node is a "father" if it is a claim and a "son" if it is evidence for that claim.

Transmission Strategies

Speakers typically follow specific patterns when ordering their thoughts:

  • Pre-order: The claim is stated first, followed by its evidence.
  • Post-order: The evidence is presented first, leading up to a claim.

To handle real-world complexity, a "Hybrid" algorithm was introduced. It maintains a pointer to the last processed node and uses a stack-like mechanism to manage propositions whose logical place is not yet certain.

Figure 6: An example of mapping natural language propositions to a logical tree structure using linguistic clues like "In sum."

The Evidence Oracle and Linguistic Clues

To focus on structure, the research assumes an "Evidence Oracle"—a semantic component that determines if Proposition A serves as evidence for Proposition B. Furthermore, linguistic connectives (e.g., "In addition," "As a result") are categorized as functional instructions that guide the system on where to attach new propositions in the tree.

Key Insights and Algorithmic Efficiency

A pivotal insight is the "Right Border" constraint. In a coherent argument, a new proposition is almost always related to a node on the "right border" of the current tree (the path from the last node up to the root). This drastically prunes the search space.

Figure 1: Illustration of a tree structure where a new node (8) is attached, demonstrating the search for logical neighbors on the right border.

The research provides a formal proof that the Hybrid algorithm runs in linear time ($O(n)$) relative to the number of propositions. This is a significant improvement over $O(n^2)$ approaches, demonstrating that following speaker expectations allows for efficient reconstruction.

Taxonomy of Connectives

The author successfully mapped six categories of linguistic connectives to specific tree-building operations:

  1. Parallel: Locates a "brother" node (e.g., "Secondly").
  2. Inference: Identifies the current node as a "son" of the previous one.
  3. Detail: Identifies the current node as a "father" (claim) to the previous one.
  4. Summary: Signals the system to pick up multiple "sons" at the same level (e.g., "In sum").
  5. Reformulation: A unique case where a node is both "son" and "father" (equivalence).
  6. Contrast: Handles "counter-evidence" or "counter-claims."

Limitations and Conclusion

While groundbreaking, the model relies on the "Oracle Problem"—the assumption that a system can easily determine evidence relationships. Additionally, the strict tree structure does not account for Directed Acyclic Graphs (DAGs), where one piece of evidence might support multiple claims. Despite these limitations, the work remains a cornerstone in understanding the logic of discourse independently of world knowledge.

Summary updated

BibTeX
@inproceedings{cohen1981investigation,
  author = {Robin Cohen},
  title = {{Investigation of Processing Strategies for the Structural Analysis of Arguments}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1981},
  doi = {10.3115/981923.981943},
}
Figure