Syntactic Analysis: Literature seminars

During the literature seminars we will discuss the given scientific articles, and questions related to them. You are supposed to prepare according to the guidelines below. The seminars are meant as learning opportunities, where you will get help to understand two scientific articles related to the course in detail. Both articles present methods for using neural networks in parsing, thus complementing the lectures.

We are aware that you did not yet take the machine learning course. You may thus view the neural networks mainly as black boxes, and do not have to understand that part, including the maths, in detail.

Groups

The seminars will be held in smaller groups.

Preparation

For both seminars you are expected to read the assigned article carefully. You should also prepare answers to the questions and discussion points below and bring them to the seminar.
  • Seminar 1: Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, Noah A. Smith. Recurrent Neural Network Grammars. NAACL 2016. (You do not need to go into any detail about the math, just strive to understand the higher-level reasoning behind it.)
    • The seminar will be held in small groups in 2-0025, Wednesday February 7, 9-11, 2024.
      • Group 1, 09.15 - 10.00: Astrid, Yiyang, Anneth, Nilay, Lucie, Jiati, Xiaotian, Fei
      • Group 2, 10.15 - 11.00: Hanna, Marie, Stribor, Martina, Irene, Xingming, Irem, Yixing
    • Questions and discussion points:
      1. Can you briefly summarize the article (in about a minute)?
      2. What do you think is the most important point the authors make in the article?
      3. What do YOU personally think is the most interesting point in the article?
      4. What are the main differences between RNNG and CKY?
      5. What are the similarities/differences between RNNG parsing and transition-based dependency parsing?
      6. Why is the time complexity of RNNGs lower than for CKY?
      7. What did you learn when reading the article? (Or: why do you think you did not learn anything?)
      8. Is there something you do not understand or that is difficult to understand?
      9. Come up with some further questions/issues to discuss!
      10. Would you recommend this article? Why or why not?
  • Seminar 2: Eliyahu Kiperwasser and Yoav Goldberg. Simple and Accurate Dependency Parsing Using Bidirectional LSTM Feature Representations. TACL. Volume 4, 2016
    • The seminar will be held in small groups in 2-2025, Wednesday March 4, 13-16, 2024:
      • Group 1, 13.15–14.00: Nilay, Xiaotian, Yixing, Martina, Stribor, Astrid, Yiyang, Irem
      • Group 2, 14.15–15.00: Marie, Anneth, Lucie, Fei, Xingming, Hanna, Irene, Jiati
    • Questions:
        1. Can you briefly summarize the article (in about a minute)?
        2. What do you think is the most important point the authors make in the article?
        3. What do YOU personally think is the most interesting point in the article?
        4. Compare graph-based and transition-based dependency parsing. What are the main strengths and weaknesses of each?
        5. Tables 1 and 3 shows the results both for the K&G parser and for other parsers. Try to see what the impact seems to be of techniques like: external word embeddings, beam search, dyanmic oracles, and POS-tags (gold or predicted) by comparing scores for systems with different variants with respect to each technique.
        6. What did you learn when reading the article? (Or: why do you think you did not learn anything?)
        7. Is there something you do not understand or that is difficult to understand?
        8. Come up with some further questions/issues to discuss!
        9. Would you recommend this article? Why or why not?

Grading

The two seminars are obligatory. In order to get a passing grade on the seminars you need to read the articles, prepare according to the instructions, and be active during the seminars. All students should be prepared to answer all questions! It should be clear that you have read and analysed the article, and prepared the questions, but it is perfectly fine if you found the articles difficult, did not understand everything, or if you have misunderstood some parts. Part of the purpose of the seminars are to help you in understanding these articles.

Missing a seminar

If you miss a seminar, or if you are unprepared, you will have to do a complementary task, consisting of a written report where you briefly summarize the article, and discuss the questions/issues given above. The report should be roughly 1.5-2 pages per seminar. The report should be handed in by email to Sara, and you should also book an appointment with Sara, to briefly discuss your report. The deadlines for such reports are:
  • Seminar 1: February 14
  • Seminar 2: March 11
  • Backup deadline for both seminars: April 15
Note that these deadlines only concern students who do not get a passing grade by attending and being active at the seminar sessions.