Aligning Articles in TV Newscasts and Newspapers

Yasuhiko Watanabe, Yoshihiro Okada, Kengo Kaneji, Makoto NagaoCOLING 1998

TL;DR

This paper presents a novel method for automatically aligning articles from TV newscasts and newspapers, enabling more effective cross-media information retrieval. By leveraging textual information (telops) from TV broadcasts and structural cues from newspaper articles, the system achieves high accuracy in linking related content, paving the way for integrated browsing and search experiences.

What Problem Does This Paper Address?

In today's media landscape, information is disseminated through diverse channels, with TV newscasts and newspapers being two prominent examples. While both cover similar events, their presentation styles and information structures differ significantly. TV newscasts offer intuitive, visual, and auditory reporting, but precise details can be elusive. Newspapers, conversely, provide detailed textual accounts but lack the immediacy and visual impact of broadcast media. To effectively utilize both sources, a crucial challenge is to establish correspondences between related articles. Without this alignment, users cannot easily navigate or retrieve information across these distinct media formats, hindering a comprehensive understanding of news events. This paper tackles this challenge by proposing a method to automatically align TV newscast segments with corresponding newspaper articles.

Methods

The core of this work lies in two main components: extracting relevant information from TV newscasts and analyzing the structure of newspaper articles to facilitate alignment.

For TV newscasts, the authors focus on extracting textual information from "telops" (on-screen text overlays). They propose a multi-step process:

  1. Manual extraction of telop text: Initially, text is extracted from TV images.
  2. Text segmentation: Extracted text is divided into meaningful segments based on character size and spacing changes.
  3. Morphological analysis: A Japanese morphological analyzer (JUMAN) is used to process the segmented text.
  4. Telop analysis: Telops are categorized based on their function (e.g., scene explanation, fact explanation, quotation).
  5. Noun extraction: Crucially, nouns are extracted only from telops that explain the content of TV images or state facts, excluding direct quotes or non-report-related information. This is because these specific types of telops are deemed to contain the most relevant keywords for alignment.

For newspaper articles, the authors exploit their structured nature, particularly when available in HTML format. They divide articles into four key parts:

  1. Headline: The main title of the article.
  2. Explanation of pictures: Text accompanying images.
  3. First paragraph: The introductory section.
  4. The rest: The remaining body of the article.

The alignment process itself involves calculating reliability scores between a TV newscast article and potential newspaper article matches. These scores are based on:

  • Location of keywords: Where words appear within each article type.
  • Frequency of keywords: How often words appear.
  • Length of keywords: The length of the words themselves.

A weighted scoring function, SCORE(x, y), is defined, incorporating weights w(i, j) that reflect the importance of word locations in both newspaper sections and TV newscast segments. The newspaper article with the highest reliability score above a certain threshold is considered aligned.

Datasets

The paper does not explicitly mention the creation or use of specific training datasets. The evaluation was performed using:

  • 143 articles from the NHK evening TV newscasts.
  • Corresponding articles from the Asahi newspaper (obtained from the Internet).

The alignment was performed between TV newscasts and newspapers from the same day, as well as between TV newscasts and the morning paper of the following day.

Key Insights and Ideas

  • Telops as a rich source of information: The paper highlights that even secondary textual elements like telops in TV newscasts can contain crucial keywords for understanding and aligning content with other media.
  • Structural cues in newspapers are vital: The distinct sections of a newspaper article (headline, first paragraph, etc.) provide valuable structural information that can be leveraged for matching.
  • Keyword distribution is uneven: The authors correctly identify that keywords are not uniformly distributed across articles; they are concentrated in specific sections, which informs their weighting strategy.
  • Cross-media alignment enables new functionalities: The successful alignment opens doors for integrated browsing and retrieval systems, allowing users to seamlessly navigate between TV and newspaper content.

Key Results

The proposed alignment method achieved impressive results:

  • Recall: 97%
  • Precision: 89%

These figures indicate a highly effective system for linking related articles across TV newscasts and newspapers. The authors also developed a browsing and retrieval system that utilizes these alignment results, allowing users to search and access content based on dates and full-text queries, with cross-referencing between media types.

Limitations and Future Work

The authors acknowledge a key limitation: the abbreviation of words. For instance, "shinyo-kinko" (credit association) being abbreviated to "shinkin" can lower reliability scores.

To address this, they propose future work focusing on:

  • Improving alignment performance using dynamic programming matching: This technique is suggested to handle string variations and abbreviations more effectively.
  • Developing specialized alignment methods for specific content types: The authors note that certain TV news categories, like sports or weather, have fixed styles and may require different alignment approaches.

The paper also implicitly suggests that further refinement of keyword extraction and weighting strategies could enhance performance.

Summary updated

BibTeX
@inproceedings{watanabe1998aligning,
  author = {Yasuhiko Watanabe and Yoshihiro Okada and Kengo Kaneji and Makoto Nagao},
  title = {{Aligning Articles in TV Newscasts and Newspapers}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1998},
}
Figure