Abstract
This paper proposes a fundamental shift in the architecture of question-answering (QA) systems by decoupling the retrieval of data from the selection of communicative content. Traditional systems in the early 1980s often treated QA as a single-step search for a "best" answer, frequently resulting in responses that were either redundant or conversationally irrelevant. By bifurcating the process into answer formation—the identification of relevant causal chains—and answer expression—the pruning of those chains based on conversational norms—the proposed system generates responses that are more natural, informative, and aligned with human expectations.
Introduction
In the early 1980s, the primary challenge for question-answering systems was not just finding information, but determining which information was worth sharing. Most contemporary systems viewed the task as a selection process: gather a list of potential answers and pick the most direct one. However, this approach struggled with complex narratives and "why" questions.
If a system simply identifies the end of a causal chain, it often produces "normative" facts—standard behaviors that are technically true but conversationally useless. For instance, if a character buys a book to study for an exam, a simple system might state she bought it "to have it." While accurate, such a response misses the point of the user's query. The challenge addressed by Marc Luria is determining which parts of a complex knowledge structure are actually "interesting" or "important" to a human listener.
Methodology
Luria introduces a two-stage pipeline integrated with story-understanding tools such as PHRAN (for parsing) and FAUSTUS (for goal-based analysis). The process is divided into two distinct components:
1. The Answer-Formulator
This component is responsible for searching a database of "causal chains"—sequences of events, inferences, and goals generated during the initial story-understanding phase. It identifies the entire relevant path in the database that connects the user's question to the eventual outcome of the narrative.
2. The Intelligent-Expresser
This is the core innovation of the paper. The Intelligent-Expresser takes the raw causal chain provided by the Formulator and applies "general rules of expression" to filter it. A primary rule is the avoidance of redundancy: the system should not inform the user of things they already know. It identifies "normative" steps (e.g., the fact that one must buy a book before reading it) and skips them. Instead, it focuses on "important" inferences, such as specific high-level goals (e.g., passing a math test) or abnormal events (e.g., a character attempting to learn via osmosis).
The resulting pruned conceptual form is then passed to a natural language generator, such as PHRED, to produce the final response.
Evaluation and Case Studies
The system was evaluated using "simple stories" designed to test goal-based reasoning rather than large-scale datasets. Two primary scenarios were utilized:
- The Susan Story: A narrative where a character named Susan saves money, travels to a bookstore, and buys a recommended book to help her succeed on a math test. This tests the system's ability to navigate standard goal-oriented behavior.
- The Osmosis Variant: A modified version where Susan attempts to learn by putting the book on her head. This variant tests how the system handles "abnormal" or "important" inferences that deviate from standard human behavior.
Key Insights
Answers as Subsets of Chains
A central thesis of the paper is that an answer is not a single data point but a strategically pruned subset of a causal path. By viewing the answer as a path rather than a destination, the system can provide context that a single-point retrieval system would miss.
The Normative Filter
Humans naturally assume standard sub-goals. If an AI mentions every mundane step in a plan (e.g., "She opened the door, she walked in, she saw the book..."), it becomes tedious. By identifying and skipping these "normative purposes," the AI mimics human conversational economy.
Handling Yes/No Questions
The paper argues that a simple "Yes" or "No" is often an insufficient answer. The two-stage process allows the system to provide a "Yes" followed by a helpful elaboration by finding the closest conceptual match in the causal chain (e.g., "Yes, she rode her bike there").
Results
The system demonstrated qualitative improvements over contemporary algorithms, such as those developed by Winograd or Lehnert. Key results include:
- Reduced Redundancy: The system successfully avoided stating obvious steps in a plan.
- Improved "Why" Responses: For the question "Why did Susan buy the book?", the system correctly prioritized the goal of "doing well on her math test" over the mundane fact of "wanting to have the book."
- Prioritization of the Unusual: In the "osmosis" example, the system correctly identified that the unusual method of learning was the most important information to express, as it deviated most significantly from the user's expectations.
Limitations and Future Work
At the time of publication, the program was not yet fully integrated with the PHRED natural language generator; examples were translated from the system's internal conceptual forms. Furthermore, the system relies heavily on the quality of the underlying causal chains provided by the story-understander (FAUSTUS). Luria suggests that this "intelligent expression" mechanism could eventually be expanded beyond question answering to general conversation and automated summarization tasks.
Summary updated