The DiaBiz-EN Corpus: Evaluating Translated Special-domain Spoken Language

The paper presents DiaBiz-EN, a parallel corpus of transcriptions of phone-based customer-agent interactions in nine business domains derived from the Polish DiaBiz corpus of transcribed call-centre dialogues (Pęzik et al. 2022). While DiaBiz-EN is a valuable resource for variety of studies of special-domain conversational language, in this paper we focus on evaluating the quality of its translation component in a reception experiment conducted with two main groups of participants: L1 speakers of Polish and L1 speakers of English. In each of the conditions the dialogues were presented in the participants’ L1. Therefore, while the Polish-speaking participants engaged with dialogues as originally recorded and transcribed for the DiaBiz corpus, dialogues for the English-speaking participants were translated from Polish. With this set-up, the study’s objective is generally two-fold. First, it sheds light on facets of the Polish and English readers’ experience – cognitive load, comprehension, perception of the conversations, with emphasis on the naturalness of language employed by the interactants, and perception of the interactants. Second, the study offers a language user perspective on the construction of spoken parallel corpora. As reception data collected based on stimuli from DiaBiz and DiaBiz-EN generally converges, with comprehension as well as perceived naturalness remaining unaffected for the English users, translation is demonstrated to serve as a method for developing a reliable language resource for a number of possible applications, including development and evaluation of natural language processing systems.