Corpus IST-Erasmus

Published: 22 October 2024| Version 1 | DOI: 10.17632/nmszc6pcnc.1
Contributor:
Serap Onen

Description

Corpus IST-Erasmus is a comprehensive dataset collected for a doctoral dissertation by Önen (2014), focusing on spoken English as a Lingua Franca (ELF) interactions among Erasmus students studying in Istanbul. The dataset includes the full recordings of 54 naturally occurring speech events, comprising 29 interviews and 25 focus group meetings, totaling 10 hours and 47 minutes of interaction. In addition to the recordings, the dataset also contains raw transcriptions of these interactions, amounting to 93,913 words, as well as marked and coded transcriptions, following the VOICE convention for spoken ELF communication. The dataset represents interactions involving 79 participants from diverse linguistic and cultural backgrounds, with 24 different first languages (L1s). This rich dataset offers multiple layers of analysis and provides an invaluable resource for researchers exploring various dimensions of multilingual and multicultural communication, particularly in the context of ELF.

Files

Steps to reproduce

The Corpus IST-Erasmus was compiled for a doctoral study supported by the Scientific Research Projects Coordination Unit (BAP) of Istanbul University. This quantitative and descriptive research aimed to explore spoken ELF (English as a Lingua Franca) interactions among ELF speakers, primarily incoming Erasmus students at universities in Istanbul. The corpus consists of 54 speech events, comprising 29 interviews and 25 focus group meetings, totaling 10 hours, 47 minutes, and 26 seconds of recorded spoken interactions. The data contains 93,913 words of transcribed speech. Participants were 79 Erasmus students from 4 state and 6 foundation universities, representing 24 different first languages (L1s), such as German, Polish, Italian, Dutch, and others. The data collection process spanned three months, during which naturally occurring conversations were recorded. The corpus includes three main datasets: Audio recordings of all interactions; Verbatim transcriptions of the recorded data; Coded transcriptions of the interactions, following the VOICE (Version 2.1) transcription conventions. The corpus offers a valuable resource for examining real-life ELF communication across diverse linguistic backgrounds.

Institutions

Istanbul Universitesi, Istanbul Universitesi-Cerrahpasa

Categories

Audio Recording, English as a Lingua Franca, Corpus Linguistics, Corpus Analysis

Licence