A Text Categorization Based on a Summarization Extraction

Sue J. Ker, Jen-Nan ChenACL 2000

TL;DR

This paper proposes a novel text categorization technique that leverages summarization principles, specifically by extracting categorization knowledge solely from document titles. By combining word frequency and positional information within titles, the method achieves competitive performance on the Reuters news corpus, offering a computationally efficient solution for online document classification.

What Problem Does This Paper Address?

The explosion of online information necessitates efficient methods for categorizing documents to enable quick and accurate access. Traditional text categorization systems, while effective, can be computationally intensive. This paper addresses the challenge of developing a text categorization system that is both accurate and computationally efficient, particularly for real-time applications. The authors aim to find a balance between performance and speed by focusing on a smaller, yet informative, part of the document: the title.

Methods

The core of this research lies in a text summarization-inspired approach to text categorization. The authors propose a method that extracts categorization knowledge exclusively from document titles. This involves:

  • Preprocessing and Feature Selection: Texts are tokenized, converted to lowercase, stop words are removed, and words are stemmed to create "terms" which serve as features.
  • Term Weighting: A crucial aspect is assigning weights to these terms within each category. The paper explores several formulas for term weighting, denoted as $W(f, c)$, which combine term frequency ($tf$) and inverse document frequency ($idf$). Notably, they introduce normalization of term frequency across categories and incorporate the probability of a category. Equations 1-a through 1-d present these formulas.
  • Category Ranking: Once terms are weighted, a document is ranked against each category. The ranking score $R(c, d)$ for category $c$ and document $d$ is calculated by summing the weighted frequencies of features present in the document (Equation 4).

The novelty lies in the exclusive focus on the title field for feature extraction and the application of summarization-like weighting schemes to these title-based features.

Datasets

The experiments were conducted on the Reuters corpus. Specifically, the authors used Reuters version 3, a formatted version constructed by Apte et al. This version contains 7,789 training documents and 3,309 test documents across 93 categories. The paper notes that there are multiple versions of the Reuters corpus, and they selected this specific one for comparability with other research.

Key Insights and Ideas

  • Titles are Sufficiently Informative: The research strongly suggests that document titles alone contain enough information for effective text categorization, challenging the notion that full document content is always necessary.
  • Summarization Principles Apply to Categorization: By adapting techniques from text summarization (like word frequency and positional importance), the authors demonstrate a successful new avenue for text categorization.
  • Computational Efficiency: Focusing solely on titles significantly reduces processing time, making the system highly suitable for online, real-time applications.
  • Robustness to Sparse Data: The experiments indicate that the title-based approach is not severely affected by sparse data, a common issue in text classification.

Key Results

The paper reports promising results on the Reuters corpus:

  • Method 4 (tf×idf) Achieves Best Performance: Among the tested term-weighting formulas, Method 4, which uses $tf \times idf$, achieved the highest average 11-point precision of 82.7% in the first experiment.
  • Title-Only Performance is Competitive: The title-only approach achieved an average precision of 82.7% (for Method 4 with minimum term frequency threshold of 3), which is presented as competitive with other methods that use full texts (e.g., kNN's 93% and LLSF's 92%).
  • Minimal Impact of Term Frequency Threshold: Varying the minimum term frequency threshold (from 1 to 3) showed only minor improvements, suggesting that the information in titles is concentrated and less prone to noise from very infrequent terms.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the title-only approach is efficient, it might not capture the full nuance of documents compared to methods using entire texts.
  • The performance is presented as "acceptable," implying there's room for improvement to match state-of-the-art full-text methods.
  • Future Work:
    • Hybrid Approaches: The authors suggest combining their title-based summarization technique with other schemes for potentially more efficient and accurate categorization.
  • Extended Position-Based Analysis: Investigating a hybrid structure that considers the title along with other specific positions within the document (e.g., the first sentence of the first paragraph) is proposed as a way to enrich categorization knowledge when titles are insufficient.

Summary updated

BibTeX
@inproceedings{ker2000text,
  author = {Sue J. Ker and Jen-Nan Chen},
  title = {{A Text Categorization Based on a Summarization Extraction}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2000},
}
Figure