John R. Starr; jrs294@pitt.edu; 01/29/19
Looking at parallel colloquial-speech corpora between English and Persian, I will investigate the effects that translation has on syntactic scrambling.
I'm looking to use the Tehran English-Persian parallel corpus (TEP). The corpus can be downloaded at the following URL: http://opus.nlpl.eu/TEP.php
This corpus consists of approximately 1.22 million sentence fragments of English subtitles and their Persian translations. According to an article written about the corpus (which is cited at the URL provided), the corpus is already "clean and pre-processed with special care for the Persian part"; that being said, I have not been able to examine the Persian file yet, as my computer has been acting up. The English file is an xml file that separates the sentences by the following string; " . " The OPUS resources also have pre-compiled word alignments and other formatting niceties.
As someone interested in translational studies, I am intrigued to see how the subtitles handle syntactic scrambling. Persian is a language that allows for this linguistic process, especially when spoken colloquially. I intend on using tagged sentences to determine the word order of the sentence and then examining which syntactic patterns in English initiate scrambling in Persian. I anticipate that there will be significant scrambling caused by idiomatic expressions in English of which there are few equivalents in Persian. As noted in the comments of my original proposal of this idea, I will need to find a reliable parser for Persian. Thankfully, I have found a plethora of resources to help parse the words. They can be found at the following links:
Collection of various resources: https://github.com/mhbashari/awesome-persian-nlp-ir
Paper on ParsPer, which achieves approximately 82-87% accuracy: https://www.diva-portal.org/smash/get/diva2:881563/FULLTEXT01.pdf
Paper on MaltParser and MST Parser (trained on Uppsala Persian Dependency Treebank): https://pdfs.semanticscholar.org/2759/a770ef731544e946cf47dbb198e0f6ef153e.pdf It should be noted that I don't believe the latter two parsers here would be the ones to use, as their accuracies are too low (though it would be an interesting comparison, nonetheless)!
After testing this and determining if it is true or not, I would be interested in seeing if I could build a machine learning model that anticipates the word order for the Persian sentences based on the English model, though I am not entirely sure how to do that.