An intelligent digester for interactive text processings

K. HanakataCOLING 1980

TL;DR

This paper proposes an interactive system designed to help experts digest and represent news reports about the Japanese computer industry. It focuses on bridging the gap between current natural language processing (NLP) limitations and the nuanced understanding required by domain experts, aiming to create a "reactive knowledge database" that supports their decision-making processes.

What Problem Does This Paper Address?

The paper tackles the challenge of building intelligent systems that can effectively process textual information for expert users. While current NLP systems are promising, they often struggle with understanding the deeper, contextual meaning of semantically related sentences and tend to over-analyze details that human readers naturally filter. This makes them ill-suited for tasks like synthesizing complex information for reports, where experts need to extract purposeful meaning and integrate it into their existing knowledge base. The authors specifically target the need for systems that can support experts in understanding and representing news on the foreign computer industry, a domain requiring specialized knowledge and the ability to draw conclusions about future trends and policies.

Methods

The authors propose an interactive system that doesn't aim to fully automate NLP but rather to support expert users. Key aspects of their approach include:

  • Interactive Assimilation and Abstraction: The system facilitates a "backed-up" interaction between text and the expert's knowledge. This involves both conceptual information (the literal meaning of text) and control information (how that meaning should be processed and interpreted). Experts actively refine their understanding by identifying relations between new information and their existing knowledge.
  • Frame-like Knowledge Representation: They utilize a frame-like data structure, specifically their home-made language OBJTALK, which is object-oriented. This is chosen because expert knowledge in the computer industry is often organized around well-defined concepts and roles. OBJTALK allows for defining concepts, their properties, and methods for interaction.
  • Hierarchical Script Frames: To handle the stereotypical nature of news reports within specific topics (e.g., "Main-frame-makers"), they employ hierarchical script frames. These frames represent concepts at different levels of detail, allowing for both broad topic classification and specific subtopic descriptions.
  • Multiple Window Interface: The user interface is designed around a multiple window system. This allows for parallel display of information, parallel decision-making, and a more natural way for users to interact with and manipulate frames and their contents.

Datasets

Not discussed in the paper. The paper focuses on the approach to processing textual information rather than specific datasets for training or evaluation. The system is designed to ingest and process news reports on the foreign computer industry as they become available.

Key Insights and Ideas

  • Human-like Filtering is Crucial: The paper highlights that human readers don't process every detail of a text; they selectively focus on information relevant to their goals. Current NLP systems often fail to replicate this crucial filtering mechanism.
  • Interactive Knowledge Refinement: The system's strength lies in its interactive nature, where the expert's engagement actively shapes and refines the knowledge base. This is a departure from purely automated processing.
  • Frame-like Structures for Domain Expertise: The choice of frame-like representations is justified by the structured and conceptual nature of expert knowledge in specialized domains like the computer industry.
  • Bottom-Up Generalization: The authors acknowledge the need for methods to generalize from individual instances to create abstract concepts, reflecting how users might naturally build their knowledge.
  • Window-Based Interaction Enhances Cognition: The multiple window interface is presented as a cognitively efficient way for users to manage complex information and interact with the system.

Key Results

The paper describes a proposed system and its design principles rather than presenting quantitative results from a deployed system. The "results" are more in the form of a well-defined approach and a set of design choices aimed at achieving the stated goals. The authors demonstrate the functionality of their OBJTALK language with examples of frame definitions for concepts like "main-frame-makers" and "MBI-JPN." They also illustrate the concept of hierarchical script frames and the intended use of a multiple window interface.

Limitations and Future Work

  • Current NLP Limitations: The authors explicitly state that current NLP systems are still far from achieving the ideal goal of fully automating expert processing tasks.
  • Interface Constraints: The implementation of the multiple window system is restricted by the available display terminals, with hopes for high-resolution bit-map displays in the future.
  • Need for Tailoring: While they examine their OBJTALK language, they acknowledge the need to tailor it further for their specific approach.
  • Future Work: The paper suggests continuous elevation of the system's intelligence through ongoing design and experience. They also aim to investigate the underlying rules of expert behavior in text understanding to make the system more intelligently reactive. Further research into generalization and abstraction methods for building knowledge bases from user-defined instances is also implied.

Summary updated

BibTeX
@inproceedings{hanakata1980intelligent,
  author = {K. Hanakata},
  title = {{An intelligent digester for interactive text processings}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1980},
}
Figure