Automatic title generation for Chinese spoken documents considering the special structure of the language

Lin-Shan Lee, Shun-Chuan ChenInterspeech 2003 · DOI

TL;DR

This paper tackles the challenge of automatically generating concise titles for Chinese spoken documents. It proposes using "feature units" beyond individual words, such as character segments and syllable combinations, to better capture the nuances of the Chinese language and overcome the out-of-vocabulary problem common in speech recognition.

What Problem Does This Paper Address?

In the age of abundant online multimedia content, automatically generating titles for spoken documents is crucial for efficient browsing and retrieval. While this is a challenging task even for alphabetic languages, Chinese presents unique hurdles. Chinese is not alphabetic; it uses characters that are monosyllabic and morphemic, leading to a vast number of characters but a limited set of syllables. This structure allows for flexible word formation, including many proper nouns and domain-specific terms that might not exist in standard lexicons, leading to the "out-of-vocabulary" (OOV) problem in speech recognition. Traditional title generation methods, often relying on word-level analysis, struggle with this OOV issue and the inherent ambiguity of Chinese syllables.

Methods

The paper builds upon an adaptive K-nearest neighbor (AKNN) approach for title generation. The core innovation lies in exploring alternative "feature units" to represent document content, moving beyond individual words. These feature units include:

  • Character-level units:
    • Overlapping character segments of length N (C(N), N=1, 2, 3).
  • Character pairs separated by n characters (Pc(n), n=1, 2).
  • Syllable-level units:
    • Overlapping syllable segments of length N (S(N), N=1, 2, 3).
  • Syllable pairs separated by n syllables (Ps(n), n=1, 2).

The AKNN approach, in its generating part, first identifies K nearest training documents using a baseline method. Then, it rescores these K documents using a Naive Bayesian approach. Finally, it refines the title by replacing words in the selected training document's title that are not present in the evaluation document with high TF/IDF words from the evaluation document. The paper investigates how replacing the word-level features with these new character and syllable-based feature units impacts title generation performance.

Datasets

  • Training Corpus: 151,537 Chinese news articles with human-generated titles from the Central News Agency (CNA) in Taipei.
  • Evaluation Corpus: 210 spoken news stories broadcast by FM News98 radio in Taipei. Reference titles for these spoken documents were provided by students from the Graduate Institute of Journalism at National Taiwan University.

Key Insights and Ideas

  • The OOV problem is a major bottleneck for word-based title generation in Chinese. The flexible nature of Chinese word formation means many important terms might not be in a predefined lexicon, leading to poor recognition and thus poor title generation.
  • Syllables and characters offer a more robust alternative to words for feature extraction. Because syllables are limited and characters are morphemic, they can capture meaning even when the exact word is unknown or misrecognized.
  • Bi-character (or bi-syllabic) units are particularly effective. The observation that a large percentage of common Chinese words are composed of two characters suggests that C(2) and S(2) units are good at capturing core information.
  • Integration of different feature units can yield superior results. Combining units that capture different levels of information (e.g., C(2) for core meaning and C(3) for finer details) can improve performance by leveraging complementary strengths and mitigating weaknesses.
  • Character-level units (like C(2)) can perform comparably to words, even with lower recognition accuracy, due to their ability to bypass the OOV problem.

Key Results

The paper presents initial experimental results using F1 scores as the performance metric.

  • Baseline (words): Achieved a certain F1 score.
  • Character-level units:
    • C(1) performed almost as well as words.
  • C(2) significantly outperformed the word-based baseline.
  • C(3) performed poorly, likely due to noise.
  • Syllable-level units:
    • S(1) performed worse than C(1) due to syllable ambiguity.
  • S(2) performed comparably to C(2).
  • S(3) performed poorly.
  • Integration: Integrating S'(2) and S'(3) with an optimized weight (α=0.8) showed improvement over using either unit alone.
  • Comparison: Several integrated feature unit combinations (e.g., C(2)+C(3), S(2)+C(3), S(2)+S(3)) significantly outperformed the word-only baseline.
  • Key Phrase Extraction: Extracting key phrases and adding them to the lexicon also significantly improved performance over the baseline.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while characters and syllables can bypass the OOV problem, they might not always carry as clear a meaning as words.
  • The performance of syllable-based units can be affected by tone recognition errors, which are common in continuous speech.
  • The baseline word-based approach struggles with OOV words, which are often crucial for title generation.
  • Future Work: The paper suggests exploring the integration of different "feature units" to leverage their complementary strengths. The results in Figure 4 indicate that integrating overlapping segments of two or three characters/syllables, and potentially combining these with key phrases, holds promise for further performance improvements. The authors also highlight the need for intelligent systems that can handle the flexibility of Chinese word construction and translation.

Summary updated

BibTeX
@inproceedings{lee2003automatic,
  author = {Lin-Shan Lee and Shun-Chuan Chen},
  title = {{Automatic title generation for Chinese spoken documents considering the special structure of the language}},
  booktitle = {Interspeech},
  year = {2003},
  doi = {10.21437/eurospeech.2003-642},
}
Figure