Research On Story Segmentation Technologies For Spoken Documents

Story segmentation is a task of partitioning a stream of video,audio or text into topically coherent segments,each representing a specific topic.It serves as a precursor for subsequent tasks such as multimedia information retrieval,document summarization,topic extraction and classification.The spoken document is a common genre of multimedia data including broadcast news,lecture speeches,presentations and dialogues,etc.Story segmentation can be performed on acoustic signal or speech recognition transcripts.Compared with methods conducted on acoustics,it is easier to get topic-related semantic information from textual transcripts.Thus we focus on the story segmentation technologies based on transcripts.The story segmentation process includes two steps: feature representation and segmentation.Feature vectors with abundant topic information or a good segmenter can both greatly improve the performance of the story segmentation.Given the strong feature learning ability of neural network(NN),this dissertation studies topical text representations closely related to story segmentation and the corresponding story segmentation approaches.The contributions of this dissertation are as follows:(1)An NN-based distributed text representation in topic space.Proposes a distributed topical representation of text by using deep neural network(DNN)and long short term-recurrent neural network(LSTM-RNN)to directly predict the topic posterior probabilities.Multiple time resolution BOW,bottleneck features(BNF)and multi-task learning(MTL)strategies are also introduced to enhance the performance of story segmentation.Experiments on the topic detection and tracking(TDT2)corpus show that the proposed LSTM-RNN based topical distributed representation outperforms both the BOW baseline and the recently-proposed NN-based word and sentence vector representations.(2)A neural network–hidden Markov model(NN-HMM)based approach to story segmentation.In the proposed NN-HMM approach,each topic is represented by a hidden state which is related to an emission probability inferred from the NN predicted topic posterior probabilities.Given a text stream,a Viterbi decoder finds the hidden story sequence,with a change of topic indicating a story boundary.Experiments on TDT2 corpus show that the proposed NNHMM approach outperforms the traditional ngram based HMM approach in story segmentation task.(3)A sticky HDP-HMM(SHDP-HMM)approach to story segmentation.In the traditional HMM approach,the number of hidden states has to be known in advance.By defining an HDP prior distribution on transition matrices over countably infinite state spaces and including a parameter for self-transition bias,SHDP-HMM is able to infer the number of hidden states from the data automatically and reduce the transition probabilities among redundant topics.Experiments show that the proposed SHDP-HMM approach has better story segmentation performance than the traditional HMM approach.(4)An end-to-end story segmentation approach based on long short-term memory(LSTM)-recurrent neural network(RNN).Traditional story segmentation approaches are a two-stage pipeline consisting of feature extraction and segmentation,each has an individual objective function.Obviously,the objective function used to extract features is different from the true performance measure of story segmentation,which may degrade the segmentation results.This thesis combines the two components and optimizes them jointly,using an LSTM-RNN.Specifically,one LSTM layer is used to extract sentence vectors,and another LSTM layer is used to predict story boundaries by taking as input of the sentence vectors.The whole network is optimized directly to predict story boundaries.Experiments on TDT2 corpus show that the proposed end-to-end approach achieved comparable story segmentation results to the traditional pipeline systems.