Laplacian Eigenmaps for Automatic Story Segmentation of Broadcast News

Abstract

We propose Laplacian Eigenmaps (LE)-based approaches to automatic story segmentation on speech recognition transcripts of broadcast news. We reinforce story boundaries by applying LE analysis to sentence connective strength matrix and reveal the intrinsic geometric structure of stories. Specifically, we construct a Euclidean space in which each sentence is mapped to a vector. As a result, the original inter-sentence connective strength is reflected by the Euclidean distances between the corresponding vectors and cohesive relations between sentences become geometrically evident. Taking advantage of LE, we present three story segmentation approaches: LE-TextTiling, spectral clustering and LE-DP. In LE-DP, we formalize story segmentation as a straightforward criterion minimization problem and give a fast dynamic programming solution to it. Extensive story segmentation experiments on three corpora demonstrate that the proposed LE-based approaches achieve superior performances and significantly outperform several state-of-the-art methods. For instance, LE-TextTiling obtains a relative F1-measure increase of 17.8% on CCTV Mandarin BN corpus as compared to conventional TextTiling; LE-DP achieves a high F1-measure of 0.7460, which significantly outperforms a recent CRF-prosody approach with an F1-measure of 0.6783 on TDT2 Mandarin BN corpus.

Acoustic TextTiling for story segmentation of spoken documents

“…In this study, we take the word-based lexical TextTiling as the baseline [5]. We observe that acoustic TextTiling using the Gaussian posteriorgram representation achieves comparable F1-measures on both the short and long episodes (0.6596 and 0.3986 respectively) to word-based lexical TextTiling using LVCSR (0.6597 and 0.4197 respectively).…”

“…Our experimental results demonstrate that acoustic TextTiling perform comparably to lexical TextTiling for story segmentation of spoken documents with multiple speakers. In the future, we will extend our work to some segmentation methods in which global criteria [5] are used for story boundary identification. In this case, we will take into account all the inter-sentence acoustic similarities in a spoken document.…”

“…However, sentence delimiters are not readily available in speech. A natural thought is to use significant pauses as delimiters, resulting in pause-separated pseudo-sentences [5]. Firstly, pause-separated utterances are formed when we employ a voice activity detector (VAD) to detect significant pauses with duration longer than ψ in an audio recording.…”

“…The parameters are empirically set as ψ=0.32 second, α=10.2 seconds and β=0.96 second. It aims to prevent the pseudo-sentences from being too short or too long, because the similarity scores between two long sentences and between a long and a short sentence are probably incomparable [5].…”

“…TextTiling [4] is a typical and efficient lexical cohesion based approach that has been introduced to segment spoken documents such as broadcast news (BN) [1] and meeting recordings [3]. While TextTiling performs boundary identification using the similarities between adjacent sentences, some other methods such as dynamic programming (DP) [5] take into account some global criteria based on all the inter-sentence similarities in a spoken document. One disadvantage of the lexical cohesion based methods is that the performance heavily relies on an LVCSR system, which is built using considerable linguistic resources and requires tremendous effort to collect a large amount of training data.…”