A Study of Punctuation Prediction Model for Domain-free Text

摘要

Most automatic speech recognition systems (ASR) outputunpunctuated text. Lack of punctuation causes some problems onboth human reading and off-the-shelf natural language processingalgorithms. It is an important issue to add punctuation in therecognized text automatically. People’s communication is usuallycross many domains, each of domain has its own lexicon andwriting styles. For this reason, it is helpful for punctuationprediction with considering domain information. In previousstudies, main approaches are using acoustic feature and textualfeature, such as the part of speech tag, word vector [1], pausedurations between words [2], pitch and so on [3]. However, fewstudies take into account of individual properties of differentdomains. In this paper, we firstly testified the effects of t thedomain information on punctuation prediction task, then proposeda punctuation prediction model based on multi-task learning(MTL) approach with domain information.