Fachschaft Informatik

Prüfungsprotokolle


Prüfungsprotokolle lesen



Protokolle (2 gefunden)

Nr.PrüferFach
1031 Akbik, Alan Prof. Dr. Introduction to Natural Language Processing

Protokoll

= Datum der Prüfung
19.07.2024
= Benötigte Lernzeit als Empfehlung
1 Woche
= Verwendete Materialien (Bücher, Skripte etc...)
Vorlesungs- und Übungsfolien, Fragenkatalog

= "Atmosphäre" der Prüfung / Verhalten der Beisitzer
Entspannte Atmosphäre, man hatte am Anfang extra 10min Zeit, um die Fragen zu lesen

= Prüfungsfragen
(Nicht im Wortlaut, aber so ungefähr)

1.
1.1 Define at least 5 PoS tags.

Annotate a given sentence with
  1.2 the PoS tags from 1.1,
  1.3 a lemma for each token,
  1.4 BIO-2 tags,
  1.5 BIOES tags.

2.
2.1 Name and define the metric you would use for Part of Speech tagging (PoS).
2.2 Name and define the metric you would use for Named Entity Recognition (NER).
2.3 Why do PoS and NER require different metrics?
2.4 What metric would you use for lemmatization and why?

3.
3.1 Write pseudo code for the init method of a BiLSTM sequence labeler.
  - Assume there is already a BiLSTM class
  - Initialize the required layers, define parameters of the model
  - What are the input and output vector sizes of each layer
  - Provide explanations as comments

3.2 Write pseudo code for the forward method of the BiLSTM sequence labeler.
  - Don't use for loops, packed-padding, etc.
  - Focus on the order in which the layers from 3.1 need to be called
  - The forward method receives a token list of a sentence and should return predicted labels for that sentence

4.1 Describe a model for the task of syllable annotation (e.g. fantastic -> fan|tas|tic).
  - What is the prediction task?
  - What model architecture would you use?
  - Support your explanations with a sketch for the word "fantastic". What are inputs and outputs?

4.2 List or describe the label dictionary needed for your model.
4.3 Which metric would you use to evaluate the model and why?

5.1 Compare scalability of RNNs vs. Attention
  - Which is more scalable for large texts?
  - What (if any) is the main scalability bottleneck of RNNs?
  - What (if any) is the main scalability bottleneck of Attention?

5.2 Compare the information bottleneck of RNNs vs. Attention
  - Which will provide better predictions for long texts?
  - What (if any) is the information bottleneck for RNNs?
  - What (if any) is the information bottleneck for Attention?

5.3 Which architecture (RNN or Attention) would perform better on the task of majority detection?
  - Majority detection: Given a list of digits, predict the digit which occurs most often
  - e.g. 0 0 3 6 9 1 4 5 0: Digit 0 occurs most often 


= Fazit (Gute/schlechte Prüfung , angemessene Benotung etc...)
Gute Prüfung, Art der Fragen entsprach der der Beispielfragen
Nr.PrüferFach
1034 Akbik, Alan Prof. Dr. Introduction to Natural Language Processing

Protokoll

= Datum der Prüfung/Date of exam
08.10.2024
= Benötigte Lernzeit als Empfehlung/Required learning time

2 Wochen zum Verständnis 
2 Tage für das Kurzzeitgedächtnis
2 weeks for a good overview and understanding
2 days for short-term-memory (memorizing)

= Verwendete Materialien (Bücher, Skripte etc...)/Used Material

Folien + Übungen + aufgezeichnete VL
Slides + Exercises + recordings

= "Atmosphäre" der Prüfung - Verhalten der Beisitzer/“Atmosphere” of the examination - behavior of the assessors 

Entspannt (2. Prüfung war angeblich viel leerer als die erste)
Relaxed (2nd test was supposedly much emptier than the first)

= Prüfungsfragen/Exam questions
Task 1 PoS Tagging (18 Punkte/Points)
Sentence: "Shackled in iron chains , ..."
1.1 Define 5 PoS Tags needed for the above sentence
1.2 Tag the sentence
1.3 Regular vs universal PoS tags - what is the difference and which (do you think) has more tags and why
1.4 lexicon based PoS tagging - two reasons against using it

Task 2 New NLP Task (22)
In the following sentence we want to determine, if the used prepositions are correct, if not they should be underlined.
"Shackled by iron chains, ..." (by was underlined)
2.1 Model the problem, prediction task, input/output, architecture and draw a sketch
2.2 define the label dictionary - either write all labels or explain what is inside the dict.
2.3 How can we automatically generate training data for our model?
2.4 Evaluation metric for the model - and why

Task 3 Sentiment Analysis (22)
Sentence: "King Kong is great"
3.1 Sketch of forward pass of FastText classifier using Bigrams and the above sentence - create the bigrams, show the layers (input, one-hot encoding, embedding, mean pooling, logits, activation, output)
3.2 Pseudo-Code for Init function of FastText - Parameters, layers in correct order, length of vectors (in-/ouput) and short explanation
3.3 Pseudo-Code for predict method - order of layers, which activation funct., etc.
3.4 Bigrams vs Unigrams - difference and why are Bigrams better/necessary
3.5 Are Bigrams useful in RNNs? Yes/No and reason why

Task 4 Vanilla RNN (14)
Explain the Vanilla RNN and make a sketch when necessary
- Cell structure
- How is reccurence implemented
- input/output at time t
- functions needed
- operations needed
- learnable layers
- mathematical def. of inner workings

Task 5 RNN vs Transformer w/ attention (16)
5.1 which model is better for the following problem and why:
Input: "1 7 6 4 1 8 1 || 2"
Output: "7"
task: find the number at the position which is indicated by the first number after "||"
5.2 which model is better for simple modulo 9 calculations and why:
input: "10 + 2"
ouput: "3"

= Note (Optional)/Grade
N/A
= Fazit (Gute/schlechte Prüfung , angemessene Benotung etc...)/Summary
Gute Prüfung, alles kam dran
Nicht auf Lücke lernen (außer vielleicht beim auswendig Lernen)
Wenn man die Folien 2-3 Mal angeschaut hat und die Aufzeichnung der VL 2-3 Mal angeschaut hat, kommt man sehr gut zurecht

Good exam, everything was covered
Don't study for gaps (except perhaps when memorizing)
If you have skipped through the slides 2-3 times and watched the recording of the lecture 2-3 times, you can manage mostly everything