questions malopa - ufal/NPFL095 GitHub Wiki

Ammar, W. et al.: Many Languages, One Parser,2016

Reading section 3.1. is not required. For now, you can think of the S-LSTM as a black box. At the beginning of the class we will briefly review Transition-based dependency parsing and -if time allows it- S-LSTM.

Questions

1. What are the limitations of the mainstream approach for multilingual NLP applications? What benefits could bring a single multilingual parser?

2. Why do we need parsing? Do you know about any application where parsing is needed for the state-of-the-art results?

3. Section 3.4 mentions 3 ways of defining Language Embeddings. Preliminary experiments showed that one-hot encoding of the language ID yielded the best results. Why do you think typological properties-based embedding did not performed the best?

4. Section 3.2 mentions the usage of Fine-grained POS Tag Embeddings. If such information is only available for some languages, why using them is still beneficial?

5. What do you like and dislike about the paper? Is there anything unclear?