Application of the Direct Memory Access paradigm to natural language interlaces to knowledge-based systems

Hideto Tomabechi, Masaru TomitaCOLING 1988

TL;DR

This paper introduces DM-COMMAND, a natural language interface that leverages the Direct Memory Access (DMA) paradigm for knowledge-based systems. By integrating parsing and inference directly into a memory search, DM-COMMAND offers a more efficient and context-aware way to interact with complex systems like machine translation, proving particularly effective for handling real-world linguistic phenomena.

What Problem Does This Paper Address?

Traditional natural language interfaces (NLIs) often struggle with the dynamic and context-dependent nature of human language, especially when interacting with complex knowledge-based systems. Many existing parsers build syntactic and semantic representations that are discarded after each sentence, leading to a loss of context. Furthermore, inferential processes for tasks like disambiguation or understanding anaphora are often separate from the parsing mechanism, making them less efficient and harder to integrate. This separation can lead to slow performance or outright failure in handling nuanced language. The authors aim to bridge this gap by proposing a unified approach where understanding is a direct recognition within a memory structure, allowing for immediate action and context preservation.

Methods

The core of the paper is the application of the Direct Memory Access (DMA) paradigm to natural language interfaces. Unlike traditional parsers, DMA views understanding as a process of recognizing input sentences by matching them against pre-existing knowledge stored in memory. This memory is organized as a hierarchical network of concepts.

Key techniques include:

  • Guided Marker-Passing: This is the central mechanism. "Markers" (P-markers for prediction, A-markers for activation, and C-markers for context) are passed through the memory network. When an A-marker from an activated input concept collides with a P-marker predicting the next concept in a sequence, it signifies a successful match.
  • Concept Sequences: These represent patterns of concepts, akin to syntactic templates or phrasal lexicons, guiding the marker-passing process.
  • Concept Refinement: When a concept sequence is activated, a "concept refinement" process locates the most specific node in memory representing the input. This is crucial for instantiating general concepts to specific instances.
  • Integrated Inference: DMA inherently integrates inference. Understanding is recognition, and recognition of a request directly triggers an action. Similarly, understanding a question means knowing the answer or knowing how to derive it. This avoids separate inferential modules.
  • Multiple Semantic Networks: DM-COMMAND can access multiple, independent semantic networks. Concept sequences can guide marker-passing between these networks, allowing it to leverage both the system's domain knowledge and its own command/query knowledge.

Datasets

Not discussed in the paper. The system is described as being used for the development of a knowledge-based machine translation system at Carnegie Mellon University, implying it operates on the internal knowledge bases and grammars of that system rather than on external, pre-defined datasets.

Key Insights and Ideas

  • Understanding as Recognition: The most significant insight is reframing natural language understanding not as building representations, but as recognizing input within a rich memory structure. This fundamentally changes how context and inference are handled.
  • Seamless Integration of Parsing and Inference: By making understanding a direct memory search, DMA eliminates the need for separate inference engines. Actions or answers are triggered immediately upon successful recognition.
  • Contextual Robustness: The DMA model's ability to retain parsed information as part of the contextual knowledge in memory makes it inherently better at handling phenomena like ellipsis and anaphora, which rely heavily on prior context.
  • Portability through Separation: The separation of the core control mechanism (marker-passing) from the specific functions stored in memory nodes makes the system highly portable across different knowledge-based systems.
  • Handling Real-World Language: The paper argues that DMA is particularly well-suited for practical NLIs because real-world inputs are often short, ambiguous, and contain extra-sentential phenomena, which DMA handles more naturally than traditional parsers.

Key Results

The paper doesn't present quantitative results in terms of accuracy metrics on specific benchmarks. Instead, it focuses on the qualitative benefits and practical effectiveness of the DM-COMMAND system:

  • Effective for MT Development: DM-COMMAND has been successfully used as a development support system for researchers working on the CMU-MT system, allowing them to access and debug internal functions via natural language.
  • Dynamic Evaluation: It enables dynamic evaluation of input commands and question answering without separate inferential processes.
  • Real-time Performance: Compiled code on IBM-RT and HP9000 workstations allows for parsing and performing commanded actions virtually in real-time.
  • Handles Complex Linguistic Phenomena: The system demonstrates the capacity to handle ellipsis, anaphora, and pronoun resolution due to its context-preserving nature.

Limitations and Future Work

  • Reliance on Concept Sequences for Syntax: A current limitation is the reliance on linear concept sequences as the sole syntactic knowledge. This requires careful preparation of sequences to handle syntactically complex sentences, which can be labor-intensive and less general than external syntactic knowledge.
  • Future Work on Syntax Integration: The authors mention implementing alternative schemes for integrating phrase-structure rules into DMA, such as treating syntactic nodes as part of memory or integrating DMA into augmented context-free grammar unification. They also note that while handling syntactically complex sentences is "rather expensive," it's not impossible with careful design.
  • Portability Nuances: While highly portable, specific functions stored in memory nodes and lower-level concepts might need modification for different systems.

Summary updated

BibTeX
@inproceedings{tomabechi1988application,
  author = {Hideto Tomabechi and Masaru Tomita},
  title = {{Application of the Direct Memory Access paradigm to natural language interlaces to knowledge-based systems}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1988},
}
Figure