Maximum entropy segmentation of broadcast news

Abstract

This paper presents an automatic system for structuring and preparing a news broadcast for applications such as speech summarization, browsing, archiving and information retrieval. This process comprises transcribing the audio using an automatic speech recognizer and subsequently segmenting the text into utterances and topics. A maximum entropy approach is used to build statistical models for both utterance and topic segmentation. The experimental work addresses the effect on performance of the topic boundary detector of three factors: the types of feature used, the quality of the ASR transcripts, and the quality of the utterance boundary detector. The results show that the topic segmentation is not affected severely by transcripts errors, whereas errors in the utterance segmentation are more devastating.

Semantic passage segmentation based on sentence topics for question answering

“…Stokes and Carthy [39] proposed a lexical chaining-based approach to coarse-grained segmentation of CNN news transcripts. Christensen and Kolluru [7] presented a cascading Automatic Speech Reorganization (ASR) system with utterance and topic segmenters based on a Maximum Entropy model. Recently, Hsueh and Moore [15] investigated the problem of automatically predicating segment boundaries in spoken multiparty dialogues using a lexical cohesion-based model.…”

“…paragraphs or section boundaries [36]. The second defines passages of fixed length [5,7,25]. The last approach uses semantic clues or topicality for identifying passages [2,14,33].…”

Paragraph-based prosodic cues for speech synthesis applications

“…Linguistically oriented studies generally only examine small amounts of data, usually from a restricted domain (e.g., reading aloud of constructed examples or small domain task-oriented dialogues). However, a number of studies have successfully employed prosody and timing features for discourse segmentation using larger data sets, e.g., topic segmentation of broadcast news [26,27,28]. Unsurprisingly, features in these studies are usually based around the idea of prosodic reset and differences in pitch ranges, but quantified directly from the speech signal.…”

Automatic call section segmentation for contact-center calls

“…Christensen et al [5] presented a maximum entropy approach to find utterance and topic boundaries in news broadcasts. Similar to our approach, they use cue words and the pause length as features to recognize utterance boundaries.…”