Natural Language Processing Week 11 NPTEL Assignment Answers 2025

NPTEL Natural Language Processing Week 11 Assignment Answers 2024

1. Your teacher recommended you to read the book ‘Deep Learning with Python’. After reading the book, you want to summarize it. What kind of summarization method would you use for this purpose?

• Abstractive single document summarization
• Abstractive multi document summarization
• Extractive single document summarization
• Extractive multi document summarization

Answer: c. 1, 3

Explanation:
Since you are summarizing one book (a single document), the correct methods are single-document summarization. Depending on the approach—abstractive (paraphrasing) or extractive (copying sentences)—both are valid. Hence, 1 and 3.

---

2. Which of the following is/are True?

• a. PageRank based algorithm is used to compute the sentence centrality vector
• b. Query-focused summarization can be thought of as a simple question answering system
• c. The underlying hypothesis of LexRank algorithm is sentences that convey the theme of the document are more similar to each other
• d. All of the above

Answer: a, c

Explanation:
PageRank is indeed used in LexRank for sentence importance. LexRank assumes similar sentences represent the main theme. Query-focused summarization is related but not strictly a QA system, hence b is not strictly correct.

---

3. Which of the following is/are True?

• a. ROUGE metric is as good as human evaluation
• b. LexRank can be applied for multi-document summarization
• c. In optimization based approach for summarization, the inference problem is to select a subset S of textual units found such that summary score is maximized
• d. Maximum Marginal Relevance strives to reduce redundancy while maintaining query relevance.

Answer: b, c, d

Explanation:
ROUGE is not as good as human evaluation—it correlates, but not perfectly. LexRank works for multi-docs, optimization tries to pick the best sentence subset, and MMR balances relevance and redundancy.

---

4. It is estimated that 20% of GPT-4 generated texts are fake. Google built some AI system to filter these fake contents. An AI system claims that it can detect 99% of fake contents, and the probability for a false positive (a real content detected as fake) is 3%. Now if a content is detected as fake, then what is the probability that it is in fact a real content?

• a. 0.084
• b. 0.118
• c. 0.108
• d. None of the above

Answer: c. 0.108

Explanation:
Using Bayes’ Theorem, we calculate the false positive rate given a positive detection. The resulting probability that a detected “fake” is actually real is 10.8%.

---

5. What is the macro-averaged F1 score?

• a. 0.54
• b. 0.56
• c. 0.58
• d. 0.64

Answer: c. 0.58

Explanation:
Macro-F1 is calculated by taking the average of F1 scores across all classes equally, regardless of support (number of instances). Based on class-wise F1s, the average here is 0.58.

---

6. What is the micro-averaged precision?

• a. 0.58
• b. 0.64
• c. 0.50
• d. 0.60

Answer: d. 0.60

Explanation:
Micro-averaged precision considers the total true positives and false positives across all classes. It weights classes by their support and in this case, the calculated precision is 0.60.

---

7. What is the F1 score of boat class?

• a. 0.40
• b. 0.30
• c. 0.58
• d. 0.67

Answer: a. 0.40

Explanation:
F1 score is the harmonic mean of precision and recall. For the boat class, the low precision and/or recall values lead to an F1 of 0.40.

---

8. What is the accuracy of the classifier?

• a. 0.40
• b. 0.50
• c. 0.60
• d. 0.90

Answer: c. 0.60

Explanation:
Accuracy is calculated as (correct predictions / total predictions). Based on the confusion matrix or result data, the classifier’s accuracy is 60%.