Abstract:In natural language processing, text segmentation is a crucial task that can be approached in various ways, depending on the level of detail required. Examples include segmenting a document into different topical segments or dividing a sentence into smaller units called elementary discourse units (EDUs). Traditional methods for these tasks relied heavily on carefully crafted features, but they have limitations. SEGBOT is our proposed solution to address these limitations. SEGBOT is an all-in-one segmentation model that leverages a bidirectional recurrent neural network to encode an input text sequence. In addition, SEGBOT incorporates another recurrent neural network and a pointer network to identify text boundaries within the input sequence. Our hierarchical model can simultaneously utilize both word-level and EDU-level information for sentence-level sentiment analysis. Through our experiments, we have demonstrated that our model surpasses previous approaches on the benchmarks of Movie Review and Stanford Sentiment Treebank.