Prosodic Features for Sentence Segmentation - Dissertation Example

The most emphasis in this approach is put on the duration of pauses between words. Longer pauses are assumed to be sentence boundaries. The word boundary method presupposes that such pauses logically occur only at the end of sentences. This is true on many occasions since the place to pause is really at the end of sentences. The word boundary method is therefore quite useful especially when analyzing short sentences (Stolcke, & Shriberg, 1996, 139). The detection of sentence boundaries is one of the initial steps that lead to the understanding of speech.

The fact that speech recognizer output lacks the normal textual cues such as headers, paragraphs, sentence punctuation and capitalization was also mentioned. However, speech provides prosodic information through its durational, intonational and energy characteristics. In addition to its relevance to discourse structure in spontaneous speech and its ability to contribute to various tasks involving the extraction of information; prosodic cues are naturally unaffected by word identity. It should therefore be possible to improve the robustness of lexical information extraction methods which are based on ASR (Hakkani-Tur et al 1999).

Sentence segmentation is required for topic segmentation and is also needed to separate long stretches of audio data before parsing (Shriberg et al 2000). Sentence segmentation is critical for applications that are used for obtaining information from speech because information retrieval techniques such as machine translation, question answering and information extraction were basically developed for text based applications (Shriberg et al 2000; Cuendet et al 2007). Kolar et al (2006, p. 629) indicates that standard automatic speech recognition systems only output a raw stream of words.

It therefore means that important structural information such as punctuation is missing. Punctuation defines sentence boundaries and is fundamental to the ability of humans to understand information. Natural language processing techniques such as machine translation, information extraction and retrieval text summarization all benefit from sentence boundaries. According to Mrozinski et al (2006) spontaneous speech is generally affected negatively by ungrammatical constructions and consists of false starts, word fragments and repetitions which are representative of useless information.

Output from automatic Speech-To-Text (STT) system is affected by additional problems as the word recognition error rates in spontaneous speech is still high. Sentence segmentation can lead to an improvement in the readability and usability of such data; after which automatic speech summarization can be used to extract important data. Magimai-Doss et al (2007) indicates that the aim of sentence segmentation is the enrich the improve the unstructured word sequence output for automatic speech recognition (ASR) systems with sentence boundaries in order to make further processing by humans and machines easier.

Improvements in performance were shown in speech processing tasks such as: speech summarization, named entity extraction and part-of-speech tagging in speech, machine translation, and for aiding human readability of the output of automatic speech recognition (ASR) systems when sentence boundary information was provided. Annotation relating to sentence boundary was found to be useful in the determination of “semantically and prosodically coherent boundaries for