News-Oriented Automatic Chinese Keyword Indexing

Sujian Li, Houfeng Wang, Shiwen Yu, Chengsheng XinSIGHAN 2003 · DOI

TL;DR

This paper presents a novel system for automatically extracting keywords from Chinese news articles. It leverages linguistic techniques and a scoring system to identify meaningful phrases, overcoming challenges like the lack of explicit word boundaries in Chinese and the cost of manual indexing. The system achieves competitive results, demonstrating the effectiveness of its approach for news content.

What Problem Does This Paper Address?

In today's information-rich world, efficiently retrieving relevant information is crucial. Keywords serve as concise summaries of documents, enabling quick identification of interesting content. However, manually indexing keywords is time-consuming and expensive, especially for large volumes of text. While automatic keyword extraction is desirable, it faces challenges, particularly with languages like Chinese, which lack explicit word boundaries, making segmentation and phrase identification difficult. News articles, while abundant, often lack pre-assigned keywords, and their unique characteristics (short length, emphasis on named entities, and important information often appearing early) present both opportunities and challenges for automated systems. Existing methods for Chinese keyword extraction often rely on complex structures like PAT trees, which are computationally expensive.

Methods

The proposed system employs a three-module architecture:

  1. Recognizer Module: This module identifies candidate keywords. It uses:
  2. Frequency Statistics: Character strings appearing multiple times (threshold of 2) are considered potential keywords, as frequent terms often indicate document topics.
  3. Named Entity Recognition: It identifies named entities (person names, places, organizations, etc.) using a proper nouns lexicon and statistical/rule-based methods. These are considered important regardless of frequency.
  4. Filter Module: This module refines the candidate set by removing meaningless or redundant items using four filters:
  5. Overlapped and Dependent Items Filter: Removes substrings that are fully contained within longer strings with similar frequencies.
  6. Punctuations and Function Words Filter: Eliminates strings containing punctuation or common function words at their boundaries.
  7. Segmentation Filter: Verifies if a candidate string forms a meaningful unit within its sentence context by using a segmentation tool. If the surrounding characters are segmented into a single word with the candidate, it's deemed not an integrated unit.
  8. POS Filter: Filters out candidate strings that do not represent meaningful semantic units, such as prepositions, adverbs, adjectives, or conjunctions, and also excludes verb phrases, adjective phrases, and preposition phrases.
  9. Selector Module: This module scores the remaining candidate keywords based on various features and selects the top-scoring ones. Features include:
  10. Frequency (freq): How often the keyword appears.
  11. Length (len): The number of Chinese characters.
  12. Part of Speech (is_noun): Whether it's a noun phrase.
  13. Position (in_title, in_segl): Whether it appears in the title or first paragraph.
  14. Type (is_proper, in_sign): Whether it's a proper noun or enclosed in specific punctuation.

A scoring formula combines these features with empirically determined weights. Additionally, a content words lexicon with hierarchical relations is used to expand the keyword set by including parent terms of identified keywords, even if they don't appear in the text.

Datasets

The paper evaluates its system on 37 news articles from China Daily. These articles are categorized into:

  • 23 articles on National Politics
  • 10 articles on International Politics
  • 4 Sports News articles

The authors compare their results against manually extracted keywords from these articles and also against results from Chien's [1997] PAT-tree-based method. No new benchmark datasets are introduced.

Key Insights and Ideas

  • Leveraging News Characteristics: The system effectively exploits the inherent characteristics of news articles (short length, emphasis on named entities, early appearance of important information) to improve keyword extraction.
  • Multi-stage Filtering: The multi-stage filtering process is crucial for removing noise and ensuring that candidate keywords are meaningful linguistic units.
  • Content Word Expansion: Expanding keywords beyond those explicitly present in the text using a hierarchical lexicon is a valuable strategy for capturing broader thematic relevance.
  • Feature-based Scoring: A well-designed scoring mechanism that combines multiple linguistic and positional features allows for a nuanced evaluation of candidate keywords.
  • Balancing Precision and Recall: The authors aim for a balance, leaning towards higher recall when precision is reasonably assured, acknowledging that some "noise" might still be relevant for information retrieval.

Key Results

The system achieves an average recall of 0.523 and an average precision of 0.462 across the tested news articles.

  • For National Politics articles, recall is 0.452 and precision is 0.401.
  • For International Politics articles, recall is 0.644 and precision is 0.594.
  • For Sports News articles, recall is 0.629 and precision is 0.482.

These results are comparable to the state-of-the-art, specifically Chien's [1997] system, which achieved a precision of 0.43 but a lower recall of 0.30. The authors note that while only 46% of extracted keywords exactly match manual ones, the remaining extracted keywords are relevant. Conversely, about 52% of manual keywords are identified by their system, with many of the missed ones being proper nouns that appear only once.

Limitations and Future Work

  • Parameter Tuning: The parameters in the scoring formula are currently set by experience values due to the limited training corpus.
  • Proper Noun Recognition: The paper acknowledges that further improvements are needed in recognizing proper nouns, as many missed keywords are single-occurrence proper nouns.

Future work includes:

  • Machine Learning for Parameter Optimization: Accumulating more keyword-annotated documents will allow for the application of machine-learning methods to set scoring parameters more objectively.
  • Improving Proper Noun Recognition: Enhancing techniques for recognizing proper nouns is a key area for future development.

Summary updated

BibTeX
@inproceedings{li2003news,
  author = {Sujian Li and Houfeng Wang and Shiwen Yu and Chengsheng Xin},
  title = {{News-Oriented Automatic Chinese Keyword Indexing}},
  booktitle = {SIGHAN},
  year = {2003},
  doi = {10.3115/1119250.1119263},
}
Figure