NLP Concepts DY0-001 Practice Questions (Page 2)

Questions 6–10

1. 6expert · hard

   A legal tech startup is building a system to find similar past court cases. They have 500,000 case documents. They tried using TF-IDF with cosine similarity, but the results are poor because synonyms (e.g., 'car' vs. 'automobile') are not matched. They want to improve semantic matching without losing the ability to weight important terms. What should they do?

   • AUse only word embeddings and ignore TF-IDF
   • BApply topic modeling to each case and compare topic distributions
   • CReplace TF-IDF with raw word counts and cosine similarity
   • DCombine TF-IDF weighting with word embeddings by averaging embedding vectors weighted by TF-IDF scores
   Reveal Solution Discussion Select an answer first
2. 7application · medium

   A healthcare chatbot must understand that a user typing 'I have a fever' means the same thing as 'my temperature is high.' The team wants to represent words so that semantically similar terms are close together in a vector space. Which approach should they use?

   • ATF-IDF vectors for each document
   • BWord embeddings trained on a large medical corpus
   • CRegex-based pattern matching
   • DOne-hot encoding of the vocabulary
   Reveal Solution Discussion Select an answer first
4. 8expert · hard

   A public health agency has 50,000 news articles and wants to track the emergence of disease outbreaks. They want to identify articles that discuss specific diseases (e.g., 'influenza', 'COVID-19') and group them by related themes. They have no labeled data. Which approach should they use?

   • ATrain a supervised classifier to detect disease mentions
   • BUse TF-IDF to rank all articles by the word 'disease'
   • CUse topic modeling to discover disease-related themes, then inspect the top terms
   • DApply sentiment analysis to all articles
   Reveal Solution Discussion Select an answer first
5. 9expert · hard

   A financial firm wants to detect fraudulent messages in customer communications. They have a small labeled dataset of 500 fraudulent and 500 legitimate messages. They want to use word embeddings but are concerned that the small dataset will not train good embeddings from scratch. What should they do?

   • ATrain embeddings from scratch on the 1,000 messages
   • BUse one-hot encoding with a linear classifier
   • CUse TF-IDF vectors instead of embeddings
   • DUse pre-trained embeddings and fine-tune them on the labeled dataset
   Reveal Solution Discussion Select an answer first
6. 10foundation · easy

   What is the primary output of a topic modeling algorithm?

   • AA vector representation of each word in the vocabulary
   • BA set of topics, each represented by a distribution of words
   • CA single label assigned to each document
   • DA list of stop words to remove from the corpus
   Reveal Solution Discussion Select an answer first