DIMACS REU: Summer 2013
Project Description
Given the nature of machine translation (MT), and the constant attempts to improve MT algorithms, we need inexpensive, repeatable, and fast metrics to assess the quality of a given MT algorithm. We believe we can improve on the current standard automated MT metric, Bleu.
Mentors:
Michael Marcondes de Freitas (Mentor)
Eduardo D. Sontag (Faculty Supervisor)
Eugene Fiorini (REU Director)
Weekly Log
Week 1
Established a general idea of what we will pursue in regards to Machine Translation (MT) and considered different angles of approach to the problem at hand (though at this time, the problem isn't very clearly defined.) I am looking at the different types of MT, with Statistical and Rule-based MT being the main focus. I've also been thinking about ways we can incorporate Graph Theory/Coding Theory to analyze results/data once obtained, though it's mostly guesswork this early on.
Week 2
Comprised of a thorough literary review, this week served to establish a clearer direction for our project to head. I covered MT and Automated MT analysis, the importance of these, and learned about the shortcomings of the current standards. In this week we've defined our goal: to examine and possibly improve the Automated MT analysis process.
Week 3
Michael found great resources to use for gathering data from a past research in the form of a sizeable collection of xml files. I've decided on using python, xml, and html to gather the data since we will be using (for now) Bing's translator. I've established a server on Herkou, as well as creating a private repository. However, I have some trouble parsing the xml files, and sending them to be translated.
Week 4
I've decided to host the webapp for translating the source files locally, and we're looking into more sources to run through. I've finally worked out parsing the xml file (Beautiful Soup is a life saver), and can move on to using Bing's Translator API.
Week 5
Established a solid structure for collecting data. While continuing progress on working with Bing Translator, we're also exploring different ways to collect data using crowd sourcing.