Usage of the Speech Disfluency Detection Method for the Machine Translation of the Transcriptions of Spoken Language

Neural machine translation falls into the category of natural language processing tasks. Despite the availability of a big number of research papers that are devoted to the improvement of the quality of the machine translation of documents, the problem of the translation of the spoken language that contains the elements of the disfluency speech is still an actual task, especially for low-resource languages like the Ukrainian language. In this paper, the problem of the neural machine translation of the transcription results of the spoken language that incorporate different elements of the disfluency speech has been considered in the case of the translation from the English language to the Ukrainian language. Different methods and software libraries for the detection of the elements of disfluency speech in English texts have been analyzed. Due to the lack of open-access corpora of the speech disfluency samples, a new synthetic labeled corpus has been created. The created corpus contains both the original version of a document and its modified version according to the different types of speech disfluency: filler words (uh, ah, etc.) and phrases (you know, I mean), reparandum-repair pairs (cases when a speaker corrects himself during the speech). The experimental verification of the effectiveness of the usage of the method of disfluency speech detection for the improvement of the machine translation of the spoken language has been performed for the pair of English and Ukrainian languages. It has been shown that the current state-of-the-art neural translation models cannot produce the appropriate translation of the elements of speech disfluency, especially, in the reparandum-repair cases. The results obtained may indicate that the mentioned method of disfluency speech detection can be used for the previous processing of the transcriptions of spoken dialogues for the creation of coherent translations by the usage of the different models of neural machine translation.