Lin_ICASSP

Lin_ICASSP - LEVERAGING EVALUATION METRIC-RELATED TRAINING...

Info iconThis preview shows pages 1–2. Sign up to view the full content.

View Full Document Right Arrow Icon
LEVERAGING EVALUATION METRIC-RELATED TRAINING CRITERIA FOR SPEECH SUMMARIZATION Shih-Hsiang Lin, Yu-Mei Chang, Jia-Wen Liu, and Berlin Chen National Taiwan Normal University, Taiwan {896470017, 697470133, 697470171, berlin}@ntnu.edu.tw ABSTRACT Many of the existing machine-learning approaches to speech summarization cast important sentence selection as a two-class classification problem and have shown empirical success for a wide variety of summarization tasks. However, the imbalanced- data problem sometimes results in a trained speech summarizer with unsatisfactory performance. On the other hand, training the summarizer by improving the associated classification accuracy does not always lead to better summarization evaluation performance. In view of such phenomena, we hence investigate two different training criteria to alleviate the negative effects caused by them, as well as to boost the summarizer’s performance. One is to learn the classification capability of a summarizer on the basis of the pair-wise ordering information of sentences in a training document according to a degree of importance. The other is to train the summarizer by directly maximizing the associated evaluation score. Experimental results on the broadcast news summarization task show that these two training criteria can give substantial improvements over the baseline SVM summarization system. Index Terms— speech summarization, sentence-classification, imbalanced-data, ranking capability, evaluation metric 1. INTRODUCTION Speech summarization is anticipated to distill important information and remove redundant and incorrect information from spoken documents, enabling user to efficiently review spoken documents and understand the associated topics quickly [1-6]. A summary can be either abstractive or extractive. In abstractive summarization, a fluent and concise abstract that reflects the key concepts of a document is generated, whereas in extractive summarization, the summary is usually formed by selecting salient sentences from the original document. In this paper, we focus exclusively on extractive speech summarization, even though we will typically omit the qualifier "extractive." Aside from traditional ad-hoc summarization methods [1], such as those based on document structure, linguistic or prosodic information, and proximity or significance measures to identify salient sentences, the machine-learning approaches with supervised training have attracted much attention and been applied with good success in many summarization tasks [2-6]. In general, the summarization task is cast as a two-class (summary/non-summary) sentence-classification problem: A sentence with a set of indicative features is input to the classifier (or summarizer) and a decision is then returned from it on the basis of these features. Specifically, the problem of speech summarization can be formulated as follows: Construct a ranking model that assigns a classification score (or a posterior probability) of being in the summary class to each
Background image of page 1

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Image of page 2
This is the end of the preview. Sign up to access the rest of the document.

This note was uploaded on 03/06/2012 for the course CIS 630 taught by Professor Cis630 during the Spring '08 term at UPenn.

Page1 / 4

Lin_ICASSP - LEVERAGING EVALUATION METRIC-RELATED TRAINING...

This preview shows document pages 1 - 2. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online