Speech Summarization Without Lexical Features for Mandarin Broadcast News

Justin Jian Zhang, Pascale FungNAACL 2007 · OpenReview

Abstract

We present the first known empirical study on speech summarization without lexical features for Mandarin broadcast news. We evaluate acoustic, lexical, and structural features as predictors of summary sentences. We find that the summarizer yields good performance at an average F-measure of 0.5646 using only the combination of acoustic and structural features, which are independent of lexical features. In addition, we show that structural features are superior to lexical features and our summarizer performs surprisingly well at an average F-measure of 0.3914 using only acoustic features. These findings enable speech summarization without the stringent demands of high-accuracy speech recognition.

Introduction

Traditional speech summarization relies heavily on lexical features derived from transcripts. However, generating accurate text via Automatic Speech Recognition (ASR) is computationally expensive and prone to errors, particularly in noisy environments or with diverse accents. If the ASR transcript is inaccurate, the resulting summary quality degrades significantly. This study investigates whether Mandarin broadcast news can be effectively summarized using only "non-lexical" information, such as the inherent structure of a news broadcast and the prosody of the speakers' voices.

Methodology

The researchers formulated the summarization task as a binary classification problem: determining whether a given sentence should be included in a summary. A Support Vector Machine (SVM) classifier with a Radial Basis Function (RBF) kernel was employed to evaluate three categories of features:

  1. Acoustic/Prosodic Features (13 total): Extracted using the Praat tool, these include pitch (F0) range and slope, energy levels, sentence duration, and speaking rate.
  2. Structural Features (4 total): These describe a sentence's position within the broadcast, such as its location relative to the start of a story or the specific speaker turn (e.g., anchor vs. reporter).
  3. Lexical Features (8 total): Used for comparison, these include standard NLP metrics like sentence length, TF-IDF scores, and Named Entity density.

Dataset and Evaluation

The study utilized a portion of the 1997 Hub4 Mandarin corpus, consisting of 347 stories (4,748 sentences) from China Central Television (CCTV) International News. Reference summaries were manually annotated at 10%, 15%, and 20% length ratios. The data was split into 70% for training and 30% for testing.

Key Results and Analysis

The performance was evaluated using the F-measure across different feature combinations and summary ratios.

Table 1: Comparison of F-measure scores for various feature sets at different summary ratios (SR).

As shown in the results, the "No Text" model (Acoustic + Structural) achieved an average F-measure of 0.5646, which is nearly identical to the model using all features (0.5655). This represents a 24.46% improvement over the lead-sentence baseline (0.3200).

Insights into Feature Performance

  • Structural Dominance: In broadcast news, the position of a sentence is often more indicative of its importance than its content. Structural features alone outperformed lexical features (0.5645 vs. 0.5241).
  • The Acoustic Signal: Even without lexical or structural context, acoustic features alone achieved an F-measure of 0.3914, significantly outperforming the baseline.
  • Lexical Redundancy: Adding lexical data to a model already utilizing acoustic and structural features provided a negligible improvement of only 0.09%. This suggests that for this genre, prosodic and structural cues fully capture the significance of the spoken content.
Figure 1: Precision, Recall, and F-measure trends across different feature set combinations.

Conclusion

This research demonstrates that high-quality Mandarin broadcast news summarization is possible without relying on transcripts. By leveraging the "how" and "where" of speech, systems can bypass the need for perfect ASR. While these results are specific to the structured domain of news, they offer a promising path for summarizing speech in low-resource languages or processing massive audio archives where full transcription is cost-prohibitive.

Summary updated

Abstract

We present the first known empirical study on speech summarization without lexical features for Mandarin broadcast news. We evaluate acoustic, lexical and structural features as predictors of summary sentences. We find that the summarizer yields good performance at the average F-measure of 0.5646 even by using the combination of acoustic and structural features alone, which are independent of lexical features. In addition, we show that structural features are superior to lexical features and our summarizer performs surprisingly well at the average F-measure of 0.3914 by using only acoustic features. These findings enable us to summarize speech without placing a stringent demand on speech recognition accuracy.
BibTeX
@inproceedings{zhang2007speech,
  author = {Justin Jian Zhang and Pascale Fung},
  title = {{Speech Summarization Without Lexical Features for Mandarin Broadcast News}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2007},
  url = {https://openreview.net/forum?id=H1bkpfWOWr},
}
Figure