Automatic Text Summarization Based on the Global Document Annotation

Katashi Nagao, Koiti HasidaACL 1998

TL;DR

This paper introduces the Global Document Annotation (GDA) project, a novel XML-based tag set designed to imbue WWW documents with semantic and pragmatic structure. Leveraging this structured data, they propose a domain- and style-free automatic text summarization algorithm that uses spreading activation on an intra-document network, allowing for personalized summaries that reflect reader interests.

What Problem Does This Paper Address?

The World Wide Web has democratized content creation, but this has led to a deluge of information that is difficult for machines to process effectively. Traditional HTML tags primarily focus on document layout, offering little semantic information. Hyperlinks provide some navigational aid but don't inherently convey content meaning. Furthermore, the casual nature of online authoring often results in less readable content compared to traditional print media, lacking systematic quality control. This makes automatic processing, such as summarization, challenging. The authors aim to bridge this gap by creating a system that allows machines to understand the underlying structure of documents, enabling more intelligent applications like summarization.

Methods

The core of their approach is the Global Document Annotation (GDA) project, which defines a new XML-based tag set. This tag set aims to capture semantic and pragmatic structures within documents, going beyond HTML's limitations. Key features of the GDA tag set include:

  • Thematic/Rhetorical Relations (rel attribute): This attribute encodes relationships between document elements, such as thematic roles (agent, patient) or rhetorical relations (cause, concession). This allows for a deeper understanding of how different parts of a document connect.
  • Anaphora and Coreference (id, ana, crf, cotyp attributes): These attributes help track references between words and phrases, enabling the system to understand pronoun usage and identify when different expressions refer to the same entity.

Based on this GDA-tagged data, the authors propose a summarization algorithm that constructs an "intra-document network." This network represents document elements (sentences, phrases, words) as nodes and their semantic/rhetorical relationships as links. The summarization process then employs a spreading activation technique on this network. Elements with higher activation scores are selected for the summary. A key innovation is the ability to personalize summaries by biasing the activation process based on user-defined interests or by observing user browsing behavior. This allows summaries to be tailored to individual readers, a significant improvement over traditional methods that rely on superficial clues.

Datasets

The paper mentions using "online Japanese articles with GDA tags" for testing their summarization method. They also provide an example using a Wall Street Journal article to illustrate the summarization process and the summary browser. No specific benchmark datasets are introduced, but the GDA tag set itself can be seen as a framework for creating such datasets.

Key Insights and Ideas

  • Semantic Structure is Key: The paper strongly advocates for the importance of explicitly annotating semantic and pragmatic relationships within documents to enable truly intelligent machine processing.
  • Intra-Document Network for Summarization: Representing a document as a network of interconnected elements provides a rich structure for algorithms like spreading activation to identify important content.
  • Personalization is Crucial for User Experience: The ability to tailor summaries to individual user interests significantly enhances their utility and relevance, moving beyond generic summaries.
  • Domain/Style-Free Algorithm: The GDA-based summarization method is designed to be independent of specific document domains or writing styles, making it broadly applicable.

Key Results

The paper describes the development of a summarization system and a summary browser. While specific quantitative results (e.g., ROUGE scores) are not presented, the authors state that their method is "flexible enough to dynamically generate summaries of various sizes" and that "coreferences and rhetorical relations help summarization very much." They also note that their approach allows for trimming sentences and generating summaries of different lengths by adjusting the "window size" in their browser. The summary browser itself is presented as a functional tool that allows users to view original documents and their generated summaries, with resizable frames and interactive selection of text elements for customization.

Limitations and Future Work

The authors acknowledge that the "appropriate granularity of classification" for thematic roles and rhetorical relations is still an open question and will be determined by the current level of technology. They also mention that an exhaustive listing of all possible thematic roles and rhetorical relations is likely impossible.

Future work includes:

  • Extending to Hypertext Summarization: Summarizing documents that heavily rely on linked content.
  • Constructing a Large-Scale GDA Corpus: Building a substantial collection of GDA-tagged documents.
  • System Evaluation by Open Experimentation: Conducting broader tests to assess the system's performance.
  • Interactive and Intelligent Presentation: Further developing presentation methods that leverage GDA, potentially revolutionizing global and intercultural communication.
  • Making GDA Tools Publicly Available: Releasing tagging editors and browsers to encourage wider adoption.

Summary updated

BibTeX
@inproceedings{nagao1998automatic,
  author = {Katashi Nagao and Koiti Hasida},
  title = {{Automatic Text Summarization Based on the Global Document Annotation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1998},
}
Figure