Abstract
This paper explores the automated discovery of topically coherent segments within speech or text sequences. We introduce two new discriminative topic segmentation algorithms that utilize a novel measure of text similarity based on word co-occurrence statistics. By moving away from traditional generative models and employing "one-class" support vector machines, our approach achieves state-of-the-art results that are particularly robust to the noise inherent in automatic speech recognition (ASR) errors.
Introduction
Topic segmentation—the task of identifying where one topic ends and another begins in a continuous stream of data—is a foundational challenge in Natural Language Processing (NLP). It is essential for navigating large audio archives, improving speech-to-text quality with topic-specific models, and building intelligent dialogue systems.
Existing approaches often struggle with three major hurdles:
- Noise and Uncertainty: ASR is rarely perfect; errors and filler speech can confuse algorithms looking for topical shifts.
- Online Processing: Many methods require global optimization over a whole document, making them unsuitable for real-time applications like live news broadcasts.
- Similarity Measures: Standard word-counting (e.g., cosine similarity) fails to account for synonyms or semantically related terms.
Methodology
We propose two discriminative algorithms that make local decisions based on sliding windows, enabling online streaming applications. Central to both is a similarity kernel, $K_{norm}$, based on Pointwise Mutual Information (PMI). This measure uses word co-occurrence statistics from a large training corpus to understand that even if two windows do not share identical words, they may be topically related if their words frequently appear in similar contexts.
Support Vector Topic Segmentation (SV)
The more robust of our two approaches is designed specifically to handle ASR noise. It utilizes Support Vector Data Description (SVDD), essentially a "one-class SVM." For each window of text, the algorithm finds the smallest possible sphere in a high-dimensional feature space that encloses the "true" topical content while excluding outliers like ASR errors or filler words.
To identify a topic boundary, the algorithm measures the geometric distance between the spheres of two adjacent windows. A large distance (a local maximum) indicates a significant topic shift. Furthermore, we integrate word lattices from the speech recognizer to calculate "expected counts" and confidence scores, allowing the algorithm to explicitly account for recognizer uncertainty.
Experimental Setup
We evaluated our algorithms using the TDT (Topic Detection and Tracking) corpus, which includes approximately 311 hours of news recordings and over 1,300 news streams. Performance was measured using Co-occurrence Agreement Probability (CoAP) and Topic Closeness Measure (TCM), where higher percentages indicate better alignment with human-labeled boundaries.
Key Results
Our experiments demonstrate that discriminative methods focusing specifically on boundaries are more accurate and less computationally intensive than generative models like Hidden Topic Markov Models (HTMM).
As shown in the results table, the SV algorithm consistently outperformed both the simpler similarity-based (Sim) approach and the generative HTMM baseline across all input types. On clean text, SV achieved a 76.6% CoAP. On speech transcripts, the SV algorithm remained the leader, and incorporating confidence scores from ASR lattices further pushed performance to 69.2%. Compared to HTMM, the SV approach reduced relative segmentation error by approximately 29% on text and 10% on speech.
Conclusion
By modeling topical content as compact geometric descriptions, our discriminative approach provides a robust solution for topic segmentation in noisy environments. While these algorithms currently rely on local decisions to support online processing, future work could integrate these local distance measures into global graph-based frameworks to capture long-range dependencies in offline tasks.
Summary updated