Objective 2.4 Apply Problem-Specific Transformations to Data Sets CERTIFIED-DATA-SCIENCE-PRACTITIONER Practice Questions (Page 1)

Questions 1–5

1. 1expert · hard

   A data science team is building a multilingual sentiment analysis system. They have text in English, Spanish, and French. They want to tokenize the text so that contractions (e.g., 'don't' -> 'do' and 'not') are split, and they want to handle accented characters (e.g., 'café' should be tokenized as 'cafe' after normalization). They also want to remove stop words in all three languages. Which preprocessing pipeline should they use?

   • AUse a language-specific tokenizer for each language, apply accent normalization, and remove stop words using language-specific lists.
   • BUse a single generic tokenizer that splits on whitespace and punctuation, and remove stop words using an English list.
   • CUse a subword tokenizer (e.g., BPE) trained on all languages, and remove stop words using an English list.
   • DApply TF-IDF directly on the raw text without tokenization.
   Reveal Solution Discussion Select an answer first
2. 2application · medium

   A data scientist is working on a document retrieval system for a legal firm. They have a collection of 100,000 legal documents and want to rank documents by relevance to a user's query. They want to weight terms that are frequent in a document but rare across the collection more heavily. Which transformation should they apply to the document-term matrix?

   • ATF-IDF weighting
   • BTerm frequency (TF) only
   • CWord2Vec embeddings of the document
   • DBinary term presence (0/1)
   Reveal Solution Discussion Select an answer first
4. 3foundation · easy

   Which of the following best describes how Word2Vec learns word embeddings?

   • ABy using global word co-occurrence statistics across the entire corpus
   • BBy counting the frequency of each word in the corpus
   • CBy assigning a random vector to each word and never updating it
   • DBy predicting a word from its surrounding context or vice versa
   Reveal Solution Discussion Select an answer first
5. 4application · medium

   A research team is analyzing a large historical corpus of scientific papers. They want to create word embeddings that capture global statistical relationships across the entire corpus, such as the analogy 'king - man + woman = queen'. They have limited computational resources and prefer a method that uses global co-occurrence statistics rather than a neural network trained on local context windows. Which approach should they choose?

   • AApply TF-IDF and then perform SVD on the resulting matrix.
   • BUse one-hot encoding and compute cosine similarity between word vectors.
   • CTrain a GloVe model on the corpus.
   • DTrain a Word2Vec (skip-gram) model on the corpus.
   Reveal Solution Discussion Select an answer first
6. 5foundation · easy

   What does the term frequency (TF) component of TF-IDF measure?

   • AThe inverse of the number of documents containing the word
   • BThe number of documents in the corpus that contain the word
   • CHow often a word appears in a specific document
   • DThe total number of unique words in the entire corpus
   Reveal Solution Discussion Select an answer first