Topic segmentation in ASR transcripts using bidirectional RNNS for change detection

Abstract

Topic segmentation methods are mostly based on the idea of lexical cohesion, in which lexical distributions are analysed across the document and segment boundaries are marked in areas of low cohesion. We propose a novel approach for topic segmentation in speech recognition transcripts by measuring lexical cohesion using bidirectional Recurrent Neural Networks (RNN). The bidirectional RNNs capture context in the past and the following set of words. The past and following contexts are compared to perform topic change detection. In contrast to existing works based on sequence and discriminative models for topic segmentation, our approach does not use a segmented corpus nor (pseudo) topic labels for training. Our model is trained using news articles obtained from the internet. Evaluation on ASR transcripts of French TV broadcast news programs demonstrates the effectiveness of our proposed approach.Index Terms-topic segmentation, recurrent neural networks INTRODUCTIONThe problem of topic segmentation, to automatically breakdown a text document into topically coherent segments, has been studied for a long time. With the increase in multimedia content on the internet, there has been an interest to extend topic segmentation to audio-video documents. Multimedia documents like broadcast news programs, meeting recordings, telephone conversations and lectures commonly consist of information on more than one topic. For example, broadcast news present events related to politics, economy, sports, weather and so on. Automatic segmentation of such documents, into coherent segments, is required by several down stream tasks such as topic detection and tracking [1], summarisation, named entity extraction and for multimedia indexing and organisation [2]. * This work was performed while the author was a member of the Multispeech team

SECTOR: A Neural Model for Coherent Topic Segmentation and Classification

“…More recently, Sehikh et al (2017) utilized long short-term memory (LSTM) networks and showed that cohesion between bidirectional layers can be leveraged to predict topic changes. In contrast to our method, the authors focused on segmenting speech recognition transcripts on word level without explicit topic labels.…”

Topic-Based Measures of Conversation for Detecting Mild CognitiveImpairment

“…These context-dependent models assume that utterances with a similar semantic distribution share the same topic. More recent methods leverage the deep architectures, such as recurrent neural networks (RNNs) (Sehikh et al, 2017) and convolutional neural networks (CNNs) (Wang et al, 2016) to semantically encode the utterance into a vector space. Treating the topic segmentation as a sequence labeling problem, labels (i.e., topics) are then assigned to every utterance.…”

Hierarchical Text Segmentation for Medieval Manuscripts

“…Recently, proposed SegBot, a bidirectional RNN coupled with a pointer network that addresses both topic segmentation and EDU. Also, LSTM or CNN based approaches have been proposed, for instance through bidirectional layers (Sheikh et al, 2017), sentence embedding-based with four layers bidirectional LSTM (Koshorek et al, 2018) or through two symmetric CNN (Wang et al, 2017), etc. Finally, Arnold et al (2019) proposed Sector, the first LSTM-based architecture that combines topical (latent semantic content) and structural information (segmentation) as a mutual task.…”