Abstract
This paper introduces FRUMP (Fast Reading Understanding and Memory Program), a system developed at Yale University to process real-time news stories from a wire service. By utilizing "sketchy scripts"—top-down knowledge structures that prioritize key event information—FRUMP avoids the computational bottlenecks of exhaustive linguistic analysis. The system demonstrates the ability to summarize news stories in approximately 10 seconds and generate multi-lingual outputs, marking a significant advancement in practical natural language processing (NLP) during the late 1970s.
1. Introduction
In the late 1970s, natural language understanding was a computationally expensive task. Most contemporary systems attempted to build a complete, detailed representation of every sentence through bottom-up processing. While effective for small, controlled examples, these methods were far too slow for the high-volume, "noisy" world of real-time news.
Gerald DeJong recognized that humans do not read every word of a newspaper to understand an event; instead, they "skim" for key facts. FRUMP was designed to emulate this behavior, ignoring irrelevant details to focus on the "who, what, where, and why" of specific event types without getting bogged down in complex syntax.
2. The FRUMP System Architecture
The core innovation of FRUMP is the "sketchy script." This is a simplified version of the "scripts" developed by Roger Schank and others at Yale. While a detailed script might attempt to model every possible nuance of a situation, a sketchy script only contains "requests" (blank slots) for the most critical information. The system consists of two main components:
2.1 The Conceptual Fragment Parser
Inspired by Becker’s phrasal lexicon, this component does not require perfect grammar. Instead, it looks for phrases that satisfy the "requests" defined by the script. This allows the system to be robust against the varied and often messy language found in journalism.
2.2 The Script Applier
The Script Applier acts as the system's control mechanism. It selects the appropriate script by scanning the first paragraph of a story for "key requests." Once a script is chosen (e.g., a "vehicle accident" script), it guides the parser to find specific data points, such as the number of casualties or the location. FRUMP also features a basic memory component that checks if a new story is an update to a previously processed event by comparing extracted variables against its database.
3. Methodology and Data
FRUMP was tested using real-world data from the United Press International (UPI) news service. At the time of the paper’s writing, the system was directly connected to the Yale computer to process live news feeds. The system's performance was demonstrated on various stories, including a specific account of a plane crash in Mt. Serra, Italy.
4. Key Insights and Results
4.1 Top-Down Processing Efficiency
FRUMP proved that for specific tasks like summarization, a top-down approach (knowing what to look for before reading) is significantly more efficient than a bottom-up approach. By intentionally ignoring fine-grained details, the system became robust enough to handle real-world data.
4.2 Performance Metrics
- Speed: FRUMP could process a typical news story in approximately 10 seconds of CPU time on a DEC KA10 processor.
- Scope: At publication, the system could understand six types of news events: earthquakes, vehicle accidents, deaths of heads of state, floods, military fighting, and meetings of heads of state.
- Multi-lingual Output: Because FRUMP converted text into "conceptual dependency" (a language-agnostic internal representation), it could generate summaries in Spanish or Russian as easily as English by plugging extracted data into different "canned templates."
5. Limitations and Conclusion
The primary limitation of FRUMP was its breadth of knowledge. With only six scripts, the system was unable to process news that did not fit those specific categories. Furthermore, the summary generation relied on "canned templates," which lacked the linguistic flexibility of modern generative models. The system was also heavily dependent on the lead paragraph for classification; an unconventional lead could cause the system to fail to select the correct script.
Despite these limitations, FRUMP remains an early milestone in NLP, demonstrating that specialized, expectation-driven systems could successfully navigate the complexities of real-world information extraction.
Summary updated