TL;DR
This paper tackles the challenge of automatically segmenting Mandarin Chinese broadcast news into distinct stories using only prosodic (sound-based) features. They demonstrate that even in a tone language like Mandarin, where pitch is crucial for word meaning, acoustic cues like changes in pitch, intensity, and pauses reliably signal topic boundaries, achieving high accuracy without relying on text content.
What Problem Does This Paper Address?
Automatic topic segmentation is a crucial preprocessing step for many natural language processing tasks, including information retrieval, summarization, and anaphora resolution. While significant progress has been made for text-based sources and English audio, there's a notable gap in applying these techniques to non-English audio, especially tone languages like Mandarin Chinese. Mandarin's tonal nature, where pitch distinguishes word meanings, presents a unique challenge: can prosodic features still effectively signal discourse structure (like topic shifts) without interfering with lexical meaning? This paper aims to bridge this gap by investigating prosody-based topic segmentation specifically for Mandarin broadcast news.
Methods
The authors propose a prosody-based approach to segment Mandarin broadcast news. Their method focuses on extracting acoustic features from the audio signal, specifically:
- Pitch (f0): Measured in Hertz, smoothed with a median filter.
- Intensity: Measured in decibels, also smoothed.
- Silence: Duration of pauses.
- Duration: Word duration.
To account for speaker variability, they employ story-based normalization for pitch and intensity, approximating speaker normalization. They also consider both absolute and normalized word durations.
Crucially, they move beyond simple word-level features by incorporating local context-dependent measures. This involves calculating the change between the current word and the next word for normalized average pitch, normalized average intensity, and normalized word duration, along with the duration of the following silence.
For classification, they utilize a decision tree classifier (C4.5), chosen for its interpretability. To handle the imbalance between topic-final and non-topic-final words, they downsample their training and test sets to achieve a 50/50 split.
Datasets
The primary dataset used is the Topic Detection and Tracking (TDT) 3 (Wayne, 2000) collection of Mandarin Chinese broadcast news audio corpus. This collection comprises 3014 stories from approximately 113 hours of Voice of America (VOA) Mandarin Chinese news broadcasts, with corresponding transcriptions of about 740,000 words. The audio is stored in NIST Sphere format at 16KHz. The segment boundaries provided with this corpus serve as the gold standard labeling for evaluation.
Key Insights and Ideas
- Prosody transcends tone: The most significant insight is that prosodic cues (pitch, intensity, duration, pause) are robust indicators of topic boundaries in Mandarin Chinese, even though pitch is also used for lexical distinction. This suggests that intonational patterns for discourse structure are not solely a feature of non-tone languages.
- Pitch and intensity reduction signals finality: Words at the end of a topic segment consistently show decreased pitch and intensity compared to words within a segment. This is a strong, reliable cue.
- Duration and pause play supporting roles: Word duration increases at topic boundaries, and silence duration is also a significant feature, though its role can be language-specific (VOA Mandarin is noted for using silence distinctively at story boundaries).
- Local context is key: Features that capture the change in prosodic values between adjacent words are highly effective for boundary detection, highlighting the dynamic nature of prosodic signaling.
- Decision trees are effective and interpretable: The C4.5 decision tree classifier successfully leverages these prosodic features, providing a clear model of how these cues are combined for segmentation.
Key Results
The paper reports a decision tree classifier achieving 95.8% accuracy on a held-out test set for segmenting Mandarin broadcast news using only prosodic features. This is a substantial improvement over a random baseline of 50%.
When silence features were excluded, the classifier's accuracy dropped to 89.4%, underscoring the importance of silence in their model.
The authors also found no significant differences in classification accuracy when grouping words by their lexical tone, indicating that the prosodic cues for topic segmentation are largely independent of the specific lexical tones.
Limitations and Future Work
- Speaker Normalization Approximation: The paper acknowledges that story-based normalization is an approximation for speaker normalization, especially since speaker identification software was not available. This could be improved with actual speaker identification.
- Integration with Textual Features: While the focus is on prosody, the authors suggest integrating prosodic and textual features for potentially even better performance.
- Fine-grained Structure: The current work focuses on sequential story segmentation. Future work aims to explore more fine-grained sub-topic structure identification.
- Speaker Change Detection: Integrating speaker identification for normalization and detecting speaker changes is a planned future direction.
Summary updated