A Knowledge-based Machine-aided System for Chinese Text Abstraction

Benjamin K. Tsou, Hing-cheung Ho, Tom Bong-yeung Lai, Caesar Suen Lun, Hing-lung LinCOLING 1992

TL;DR

This paper introduces MACTAS, a novel system for automatically generating abstracts of Chinese newspaper editorials. It tackles the challenge of abstractive summarization by employing a knowledge-based approach that combines human input with an inference engine and a discourse manager, aiming to capture the core arguments and conclusions of an editorial.

What Problem Does This Paper Address?

Generating concise and informative abstracts from lengthy texts is a significant challenge in Natural Language Processing (NLP). This is particularly true for languages like Chinese, which have unique syntactic structures. The authors address the difficulty of creating fully automatic abstractive systems by proposing a "machine-aided" approach. This system aims to assist in abstract generation by leveraging a knowledge base and a structured dialog with a human informant, focusing on understanding the argumentative structure of editorials rather than just surface-level meaning.

Methods

MACTAS employs a multi-component architecture:

  • Parser: Performs linguistic analysis on informant responses, extracting syntactic and semantic information. It uses a bottom-up, data-driven deterministic parser adapted for Chinese, incorporating Marcus's look-ahead principle to handle sentence structures efficiently.
  • Knowledge Base Management and Inferencing: A domain-specific knowledge base, built from analyzing Chinese newspaper editorials, contains if-then rules designed to capture the flow of argumentation. An Inference Engine uses these rules to guide a dialog with the human informant, identifying conclusions and the reasoning chains leading to them. It simulates human dialog by avoiding redundant queries and allowing for jumps in the reasoning process.
  • Discourse Manager: Controls the interactive session with the informant, guided by a discourse model based on schemata (background, identification, argumentation, recommendation). It orchestrates a four-phase dialog to elicit information for filling these schemata, particularly the argumentation schema.
  • Text Generator: Constructs the final abstract in Chinese. It operates at three levels: discourse model (structure and content), rhetorical structure (paragraph generation using conjunctions), and single clause level (grammatically correct clause generation).

The system's core innovation lies in its human-in-the-loop approach, where a human informant guides the system through an open-ended question-answering session to understand the editorial's argumentative structure.

Datasets

Not discussed in the paper. The paper mentions that the knowledge base is derived from a "detailed analysis of newspaper editorials on the subject" and that the system is designed for "Chinese newspaper editorials" concerning the "safety of the nuclear power plant at Daya Bay."

Key Insights and Ideas

  • Human-in-the-loop for Abstractive Summarization: Recognizing the difficulty of full automation, MACTAS demonstrates the effectiveness of a collaborative approach where human input is crucial for understanding complex argumentative structures in text.
  • Knowledge-based Argumentation Elicitation: The use of an if-then rule-based knowledge base and an inference engine to systematically query a human informant about the logical flow of arguments is a novel way to extract the core message of an editorial.
  • Discourse Schemata for Abstract Structure: Employing discourse schemata (background, identification, argumentation, recommendation) provides a structured framework for organizing the information extracted from the editorial and subsequently generating a coherent abstract.
  • Pragmatic Design Choices: The authors wisely constrained the problem by focusing on a specific domain (nuclear power plant safety editorials) and a specific text type, making the task more tractable.

Key Results

The paper describes the development of a prototype system, MACTAS, and its architecture. While specific quantitative results (e.g., abstract quality scores, precision/recall) are not detailed, the paper highlights the successful implementation of a system that can generate Chinese abstracts through a human-machine dialog. The system is designed to be fully automatic only at the final text generation stage, with informant input in the understanding stage.

Limitations and Future Work

  • Limitations:
    • The current prototype can parse conjoined sentences but not complex sentences with multiple verbs.
  • The system relies on "unsophisticated informants," suggesting that the quality of input might be a factor.
  • The system's ability to handle a wide range of natural language input and generate complex sentences is limited.
  • Future Work:
    • Improve the range of natural language the system can process and generate, including complex sentences with multiple embedded verbs.
  • Refine the Discourse Manager based on an enlarged corpus of editorials.
  • Enhance the system's deductive ability and implement consistency checking to handle human-generated infelicities.

Summary updated

BibTeX
@inproceedings{tsou1992knowledge,
  author = {Benjamin K. Tsou and Hing-cheung Ho and Tom Bong-yeung Lai and Caesar Suen Lun and Hing-lung Lin},
  title = {{A Knowledge-based Machine-aided System for Chinese Text Abstraction}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1992},
}
Figure