TL;DR
This paper proposes "micro-actors" as a novel, flexible structure for representing knowledge in AI systems, particularly for question-answering. By treating both data and procedures as independent, message-passing "actors," they enable efficient, parallel processing and easier knowledge integration.
What Problem Does This Paper Address?
Building effective question-answering systems requires a robust way to represent knowledge. Traditional approaches often involve asserting facts or theorems, but integrating these can be complex and inefficient. The authors highlight the need for a structure that is both flexible enough to handle various kinds of knowledge (facts and theorems) and efficient for computation, especially in systems that might leverage parallel processing. They aim to bridge the gap between the theoretical power of actor models and practical implementation challenges.
Methods
The core contribution is the micro-actor model. Micro-actors are inspired by Hewitt's actor model but are designed for simpler, more practical implementation. Key characteristics include:
- Unified Structure: Both data and procedures are represented as micro-actors.
- Message Passing: The only form of communication between micro-actors is sending messages. This simplifies the control flow and allows for inherent parallelism.
- Procedural Embedding: Knowledge is "procedurally embedded" within micro-actors, meaning it's activated only when needed.
- Three-Part Structure: Each micro-actor has:
- Control Part: Determines the actor's action based on incoming messages and script matching.
- Script: Stores pairs of message patterns and corresponding actions, representing theorems.
- Acquaintance: A memory space for storing facts, data, attributes, and values.
- Sentence Numbering: Messages are tagged with sentence numbers to track the order of operations and distinguish results, enabling efficient retrieval of specific information (like remainders of divisions).
The authors suggest that this model is well-suited for implementation in LISP on parallel processors.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Simplicity through Unification: By making message passing the sole communication mechanism, the system achieves a high degree of simplicity and parallelism. Complex behaviors like function invocation or variable access are treated as special cases of message sending.
- Decentralized Control: The absence of a global state and the independent nature of micro-actors (except for message exchange) make debugging and justification of the system much easier.
- Flexibility in Knowledge Representation: Micro-actors can represent both dynamic procedures (like theorems) and static data (like facts), allowing for a unified knowledge base.
- Incremental Knowledge Addition: New knowledge can be added by creating new micro-actors or enhancing existing ones without requiring a complete overhaul of the system.
Key Results
The paper primarily presents a conceptual model and its structural advantages. While it doesn't detail quantitative experimental results, it illustrates the model's functionality through a diagram of message flow (Fig. 2) demonstrating how a question like "Is X divisible by 2?" is processed. The example shows how micro-actors like WARU (divide), AMARI (remainder), and DEARU (be) interact to determine the answer, highlighting the role of sentence numbering in retrieving specific results.
Limitations and Future Work
- Implementation Complexity: While the model is conceptually elegant, the authors acknowledge that realizing a truly complicated system based on actors can be challenging. They mention that pseudo-implementations of co-routines or parallel processing might reduce efficiency.
- LISP and Restricted Japanese: The implementation is described as being in LISP, with scripts potentially using restricted Japanese. This implies a specific technological context.
- Future Work: The paper implicitly suggests that implementing this micro-actor system on a parallel processor using LISP would be a natural next step to fully realize its potential for high-degree parallel processing. The authors also express gratitude to Prof. Hewitt, indicating a continued interest in actor-based AI research.
Summary updated