Errors in Crowdsourced Post Editing[Latin Thesis]

Seyed Sina Mir Arabshahi

The present thesis attempted to deal with the combination of machine translation and crowd(sourced) post-editing through two phases. In the first phase it aimed at identifying the most and the least frequent error types in raw machine translated and crowd(sourced) post-edited outputs. Moreover, it addressed the similarities and differences between the above mentioned errors in the two outputs. In order to do so, four English news texts containing an overall 1100 words, were uploaded on online platform of Google Translator Toolkit (GTT) which is supported by Google Translate, and consequently were collaboratively post-edited by eleven unprofessional translators/post-editors. Afterwards, both outputs were analyzed according to Vilar et al.’s (2006) error typology. The results showed that the ‘Incorrect Words’ and ‘Unknown Words’ were respectively the most and the least frequent types of errors in both outputs. The results also revealed a number of statistical differences and similarities along with a number of interesting points regarding the practicality of crowd(sourced) post editing. As for the second phase and objective however, the participants’ experience was taken into account. To this end, exploratory information was obtained through an online interview about their experience in the project. The collected data in general showed positive feedbacks, non-monetary incentives to volunteer and helpful suggestions to improve such initiatives.