Automatic Emphatic Information Extraction from Aligned Acoustic Data and Its Application on Sentence Compression

In this paper we address the following question: can useful Specific words can be prosodically emphasized in an utterance emphatic information be automatically extracted from by a speaker in order to draw attentions on them, the prevailing acoustic data without any manual feature extraction which can be modeled by pitch accents of words (Bolinger and be used to help improve the performance of natural 1958). Also referred as prosodic prominence, pitch accent language processing tasks such as sentence compression? is found to emphasize several semantic information in an utterance While sentence compression requires the models of a such as uncertainty, contrast, turn-taking cues and so good comprehension of the semantic context and the exact on, whose changes in an utterance can be perceived by listeners intention of the input sentence, we believe the supervision and thus convey certain kinds of emphasis (Terken of additional emphatic data can be a boost to the later, and 1991). The detection of prosodic prominence shows improvements the LSTM structures dealing with the former, which will be on different tasks, such as Text-to-Speech synthesis supported by our evidence. Meanwhile, with the Speech-To- and spoken language summarization. With most of Text alignment techniques, we present a faster approach to the detections of prosodic prominence are done by using automatically extract approximate emphatic patterns from acoustic features (acoustic durations and intensities, extremity aligned acoustic data, thus lowering the cost of manual of fundamental frequency minima and maxima), there feature extraction in emphatic words detection and prediction are also works investigating predictions of emphatic words and providing weak supervision as an auxiliary task using only lexical features (Brenier, Cer, and Jurafsky 2005; to improve sentence compression performance using LSTM Brenier 2008), which shows promising results and potential structures.