Report instructions

Your report needs to be written in LaTeX. You are required to use the ACL 2020 template which you can download from or edit directly on Overleaf. Make sure your names and student numbers are visible at the top. (Tip: you need to uncomment \aclfinalcopy). You can find some general tips about writing a research paper here, but note that you need to make your own judgment about what is appropriate for this project.

We expect you to use the following structure:

1. Introduction (~1 page, 2pts) - describe the problem, your research questions and goals, a summary of your findings and contributions. Please cite related work (models, data set) as part of your introduction here, since this is a short paper.
   • Introduce the task and the main goal
   • Clear research questions
   • Motivating the importance of the questions and explaining the expectations
   • How are these addressed or not addressed in the literature
   • What is your approach
   • Short summary of your findings
2. Background (~1/2-1 page, 1pt) - cover the main techniques (“building blocks”) used in your project (e.g. word embeddings, LSTM, Tree-LSTM) and intuitions behind them. Be accurate and concise.
   • How each technique that you use works (don’t just copy the formulas)
   • The relation between the techniques
3. Models (~1/2 page, 1pt) - Cover the models that you used.
   • The architecture of the final models (How do you use LSTM or Tree-LSTM for the sentiment classification task? What layers do you have, how do you do classification? What is your loss function?)
4. Experiments (~1/2 page, 1pt) - Describe your experimental setup. The information here should allow someone else to reproduce your experiments. Describe how you evaluate the models.
   • Explain the task and the data
   • Training the models (model, data, parameters and hyper parameters if the models, training algorithms, what supervision signals you use, etc.)
   • Evaluation (e.g. metrics)
5. Results and Analysis (~1 page, 4pt). Go over the results and analyse your findings.
   • Answer each of the research questions you raised in the introduction.
   • Plots and figures highlighting interesting patterns
   • What are the factors that make model A better than model B in task C? Investigate to prove their effect!
6. Conclusion (~1/4 page, 1pt). The main conclusions of your experiments.
   • What have you learned from you experiments? How does it relate to what is already known in the literature?
   • Were the results as expected? Any surprising results? Why?
   • Based on what you learned what would you suggest doing next?

You lose points for bad writing style (because you are asked to prepare a conference-style report).

• Improper use of the latex template (e.g. tweaked the template): -0.5
• Page limit is not respected: -0.5 for the first page, we stop reading beyond that (which will affect your grade for other criteria as well).
• Bad structure (e.g. missing important sections such as introduction and conclusion): -0.5 per section.
• Command of English: judged case by case

General Tips:

• Math notation – define each variable (either in running text, or in a pseudo-legenda after or before the equation).
• Define technical terminology you need.
• Avoid colloquial language – everything can be said in a scientific-sounding way.
• Avoid lengthy sentences, stay to the point.
• Do not spend space on “obvious” things.
• Do not go over the page limit. (We will deduct points for that.)
• The page limit is 4 pages excluding references and appendix. There is no strict limit to references and appendix. However, the report needs to remain fully self-contained: the appendix should only include content that is not necessary to understand your work. For example, preprocessing decisions, model parameters, pseudocode, sample system inputs/outputs, and other details that are necessary for the exact replication of your work can be put into the appendix. However,

An ideal report:

• Precise, scientific-sounding, technical, to the point
  • Little general “waffle”/chit-chat
• Not boring – because you don’t explain obvious things too much
• Efficient delivery of (only) the facts that we need to know to understand/reimplement
• Results visually well-presented and described with the correct priority of importance of sub-results
• Insightful analysis – speculation should connect to something interesting and not be too much; the reader “learns something new”
• No typos, no colloquialisms – well-considered language
• This normally means several re-draftings (re-orderings of information)