Abstract
OpEd is a pioneering 1986 system from the UCLA Artificial Intelligence Laboratory designed to model the cognitive structures of human argumentation. While contemporary natural language processing often relies on statistical patterns, OpEd attempts to "understand" editorials by mapping text to Argument Units (AUs)—abstract templates of reasoning and belief. By constructing a dynamic argument graph, the system can infer implicit stances and answer complex "why" questions regarding an author's logic.
Introduction
In the mid-1980s, natural language processing (NLP) excelled at understanding "scripts" for stereotypic situations, such as dining at a restaurant. However, editorials presented a significantly more complex challenge. They are not merely sequences of events but dense networks of opinions, justifications, and attacks on opposing views. OpEd was developed to bridge the gap between simple script-based understanding and the abstract nature of human argumentation.
The OpEd Architecture
To process professional prose, OpEd utilizes a multi-layered architecture that integrates linguistic processing with conceptual memory.
The system's core components include:
- Demon-Based Parser: Instead of rigid grammar rules, OpEd employs "demons"—small, specialized procedural units that handle tasks like word disambiguation and pronoun resolution in real-time.
- Argument Graph: Serving as the system's episodic memory, this graph represents beliefs, plans, and goals connected by "support" or "attack" links.
- Semantic Memory: This contains domain-specific knowledge (e.g., economics) and reasoning scripts that allow the system to fill in structural gaps in an author's logic.
Argument Units (AUs)
The central innovation of OpEd is the theory of Argument Units (AUs). These are domain-independent, abstract structures that package patterns of belief and reasoning. For example, an author might claim that a proposed plan will have the exact opposite effect of what its proponents intended.
By separating domain knowledge (economics) from argument structure (AUs), the researchers created a model capable of recognizing logical patterns regardless of the specific topic. The system can identify these units "top-down" (expecting a justification after a claim) or "bottom-up" (recognizing a pattern of conflicting goals and labeling it accordingly).
Key Insights and Results
OpEd demonstrates that argumentation is inherently abstract. One does not need to be an economist to recognize the logic of a "counter-productive plan" argument. Furthermore, the system can infer implicit beliefs; for instance, if an author expresses "disappointment" at a policy, OpEd infers that the author views that policy as "BAD."
The paper provides qualitative results showing OpEd successfully answering deep "Why" questions. When asked why Milton Friedman believes import limitations will cost jobs, OpEd traverses its internal reasoning script to explain the causal chain: fewer imports lead to fewer dollars for foreign countries, resulting in fewer American exports and, ultimately, job losses in export industries.
Limitations and Future Work
Despite its sophistication, OpEd faced significant scaling challenges. Expanding its lexicon and "demon" library was a manual, labor-intensive process. Additionally, the system relied on hand-coded knowledge and lacked the ability to learn new domains autonomously. Future research directions suggested by the authors include distinguishing between explaining facts and intentional persuasion, as well as exploring how humans naturally acquire abstract reasoning templates.
Summary updated