Literature Review (2007)

From LS2

Contents 1 Motivation 2 Ground Rules 3 Questions to Answer 4 Deadlines 5 Suggested Topics 6 Final Note

Motivation

Understanding and keeping up with the literature in a particular research area is crucial for long-term success in that area. As a ma jor portion of your grade in this course, you will integrate a large body of existing work into a coherent literature review that will be useful to you or to others who want to contribute further in the area.

Ground Rules

This literature review is an independent exercise: everything you write should be your own. You are welcome to talk with anyone you like about the literature, including the authors of papers you read and other members of the research community who know the area well—just be sure to give credit in your review to anyone who was helpful. You are expected to read whatever literature is necessary to meet the requirements of the pro ject; it is expected that this will include at least ten articles, and probably no more than thirty. It’s often hard to know how big an area is before you start an undertaking like this one, especially if you don’t know a lot about the area, so keep in mind that you might need to scale up or down as the semester progresses.

Questions to Answer

• First and foremost, what is the research problem? Can it be stated concisely and formally? Does everyone formulate it the same way? Are there substantive differences in the different ways of presenting the problem? How is performance evaluated? Is there controversy about

how it should be evaluated? What are the available datasets, and how can they be obtained? This section should be sufficient for a CS graduate student to start to tackle the problem, even if she is ignorant of all the existing approaches to solving the problem and has very little background in language technologies.

• What are the existing solutions? You should focus on empirical solutions, but you should not neglect any technique that is competitive or dominant. How different are these solutions from each other formally and algorithmically? Researchers do not always make it easy to answer

this question; papers are often written to trumpet what’s new without clarifying how much of the underlying solution is really the same. You will need to be critical and thoughtful. Your coverage should include all key work in this area. Imagine that you are trying to protect the reader from being accused of ignorance of important work.

• What are the major obstacles for future work? Is there some resource that would help bring on a major development? What hasn’t been explored yet? Are there formal problems that need to be tackled? What ideas do you have for future work on the problem? Any thoughts

on how to solve or avoid the current obstacles? Are there seemingly obvious ideas that haven’t been tried?

You need not carry out any implementation for this literature review. You are welcome, however, to present useful statistics or examples from key datasets—whatever will help the reader to understand the problem. Be creative.

Deadlines

• September 13 Write a one-page proposal describing your topic (what you know about it already) with an initial reading list of 5–10 articles. Feel free to set up a meeting with the instructor before writing your proposal. Post to your "my talk" wiki page by 5pm.
• October 1–5 Mandatory meeting to discuss your progress. Email instructor to set up a time to meet. If you send a draft the week before, you can expect some feedback.
• October 26 Draft due. Submit in pdf by email to the instructor by 5pm.
• November 21 Final version due. Submit in pdf by email to the instructor by 5pm.
• November 27, 29; December 4, 6 Oral presentations and discussion. You’ll give a 20-minute talk summarizing your review. This should be polished and “conference-quality.” We are considering an alternative presentation mode wherein several students engage in a panel discussion about an area relevant to their literature reviews.

Suggested Topics

These may be too large or too small; that’s why we have monthly checkpoints to make sure the review ends up being reasonable. Feel free to propose something else. The review should focus on either a coherent problem (question answering) or a major approach to solving a ma jor problem (syntax-based MT).

• Question answering (Wang, 2006)
• Textual entailment and paraphrase
• Coreference resolution
• Summarization
• Sentiment classification
• Natural language generation
• Morphology induction and modeling (Arguello, 2006)
• Syntax-driven machine translation (Ahmed and Hanneman, 2006)
• Syntax-based language modeling (Sun, 2006)
• Data-oriented parsing (DOP) and machine translation
• Natural language interfaces to databases
• Finite-state parsing
• Optimality Theory (Huggins-Daines, 2006)
• Bayesian topic models (Gimpel, 2006)

Final Note

It is sincerely hoped that the document you produce will be useful to you or to others in the future. If the area is closely related to your own research, you may want to recycle parts of your literature review in other reports or publications, or even your thesis. Even if the topic is not of special interest to you for future work, you are encouraged to put your report online for the use of others in the area. Consider it an act of scholarly community service.