Lecture 6

Question 1

What is Natural Language Generation (NLG)?

Answer 1

NLG is the process of transforming structured data into human-readable text, also known as Data-to-Text.

Question 2

Name three applications of NLG.

Answer 2

Structured Report Generation (e.g., BabyTalk project), Weather Reporting, and Question Answering Systems.

Question 3

What are the main types of NLG systems?

Answer 3

Classical (rule-based), Template-Based, Statistical/Neural (data-driven), and Hybrid Systems.

Question 4

What is a Template-Based NLG System?

Answer 4

A system that uses fixed text templates with slots for variable content, often used in predictable domains.

Question 5

Describe the advantage and disadvantage of Rule-Based NLG Systems.

Answer 5

Advantage: Produces highly accurate text for specific domains. Disadvantage: Requires extensive manual effort to define rules.

Question 6

What are the key components of a Classical NLG System?

Answer 6

Content determination, discourse structuring, aggregation, referring expression generation, lexical choice, realization, and fluency ranking.

Question 7

What is Content Determination in NLG?

Answer 7

The process of deciding what information to include in the generated text.

Question 8

What is Discourse Structuring in NLG?

Answer 8

Organizing the information into a coherent and logical flow within the generated text.

Question 9

Name three Decoding Strategies in NLG.

Answer 9

Greedy Sampling, Beam Search, and Top-K Sampling.

Question 10

What is Top-K Sampling?

Answer 10

A decoding method where only the top
𝑘
k most probable words are considered for generation, adding diversity.

Question 11

How does Temperature Sampling work in NLG?

Answer 11

Adjusts the probability distribution with a temperature parameter to control creativity and focus; lower values make generation more focused.

Question 12

What types of data are commonly used to train LLMs?

Answer 12

Web text from Common Crawl, Colossal Clean Crawled Corpus (C4), Wikipedia, news sites, and patents.

Question 13

Why is Data Quality important in NLG?

Answer 13

Low-quality data can introduce biases, toxicity, and unsafe content in the generated text.

Question 14

What is Prompting in the context of LLMs?

Answer 14

Using an input prompt to guide a language model to generate relevant text, sometimes known as in-context learning.

Question 15

What is the difference between Zero-Shot and Few-Shot Prompting?

Answer 15

Zero-shot prompting includes no examples, while few-shot prompting includes labeled examples to improve model performance.

Question 16

What is Instruction Tuning?

Answer 16

A finetuning phase where the model is trained on tasks framed as instructions, enhancing its ability to handle new tasks.

Question 17

What are Intrinsic Evaluation Metrics?

Answer 17

Metrics that measure model quality directly, focusing on fluency, coherence, and grammaticality of generated text.

Question 18

What is BLEU?

Answer 18

A reference-based metric used to evaluate text similarity, especially in machine translation.

Question 19

What does BERTScore evaluate in generated text?

Answer 19

It considers semantics to evaluate paraphrasing, capturing distant dependencies and word order.

Question 20

When is MAUVE used in evaluation?

Answer 20

For open-ended tasks, MAUVE measures how similar the distributions of human-written and model-generated texts are.

Question 21

What is the purpose of Human Evaluation in NLG?

Answer 21

To assess generated text based on fluency, relevance, coherence, diversity, and usefulness in task-specific contexts.