Assignment due 10/28.
What do you like most about the work you have been doing?
Personally, I enjoy reading about our assigned topic, as our problem includes aspects of linguistics and sociology (and of course I find the actual NLP topics interesting too). I feel it is beneficial to my team and myself to take notes using the paper analysis outline given to us at the beginning of the quarter, as it captures the gist of each paper in a way that is high-level while containing vital details.
What do you like least?
There is nothing so far that I actively dislike about what we have been doing. I suppose this coming week will be the hardest so far, as it's midterms week (at least for me) and we'll have to reevaluate how much time we have to spend on research versus for our other classes. I can see myself not liking that.
What is the biggest concern or question you have?
Professor Wang has been very clear about the expectations he has for us for this quarter, and although I am confident that he and Sharon will be able to help us complete our position paper on time, I am not sure our final product will be our best-ever work because of time constraints. We are going into Week 5 now and we are expected to have a complete draft completed by Week 9, both because of submission deadlines and because it would be awful to have that pressure during dead week. So in those four weeks we have to complete our literature review, run experiments, and write sections on our abstract, introduction, related works, and findings.
The main problem we are looking into is the disparity between the efficacy of natural language processing algorithms on English versus that of other languages. Current mainstream models work increasingly well on English, but when trained and used on different languages (even related ones, such as German), their effectiveness drops significantly.
Literature in this subject is lacking, as this is a recently observed phenomenon. The most recent works regarding this issue are not the most credible, either, because the sizes of their text corpora are different (meaning, there is more test data for English than for other languages, so when we train the same model using these two unequal data sets and then measure their accuracy, how much of the resulting difference is due to algorithmic problems and how much is due to the training being different?).
Most of our challenges as a group have to do the terminology and math involved in the papers we must soon start to write about. Since we will have to analyze and criticize these related works, it is imperative that we understand them thoroughly.
Group reflection: