What makes data unstructured for classification purposes?

1. It is stored outside a relational database
2. It contains missing values
3. It has no fixed schema of typed columns, and its predictive signal exists inside raw content
4. It cannot be stored digitally

Which item is an example of unstructured data?

1. A transaction table containing date, amount, and merchant columns
2. A collection of customer support call recordings
3. A spreadsheet containing monthly sales totals
4. A database table containing inventory quantities

What is metadata in the context of unstructured content?

1. The raw content itself
2. The final model prediction
3. Descriptive information such as file name, format, author, size, and timestamp
4. The loss function used during training

Why must unstructured content usually be transformed before classification?

1. Classifiers normally require numerical representations
2. Unstructured content must always be translated into English
3. Every file must be stored in a spreadsheet
4. Unstructured data cannot have target labels

Which pairing of content type and numerical representation is correct?

1. Text to spectrogram, image to word counts, audio to database rows
2. Text to token vectors, image to pixel tensors, audio to spectrograms
3. Text to GPS coordinates, image to stop words, audio to one-hot categories only
4. All content types must use exactly the same raw representation

What does feature extraction mean for unstructured data?

1. Deleting every observation containing missing information
2. Transforming raw content into informative numerical features
3. Dividing data into training and test sets
4. Removing the target labels

What is tokenization in Natural Language Processing?

1. Dividing text into units such as words, subwords, or sentences
2. Translating every document into another language
3. Removing all punctuation and numbers
4. Calculating the final classification accuracy

Why are subword tokenizers useful?

1. They guarantee perfect translation
2. They can represent unfamiliar words using smaller known units
3. They remove the need for training data
4. They convert every document into one token

What is the purpose of stop-word removal?

1. Remove every uncommon word
2. Remove selected frequently occurring words that may add little task-specific information
3. Delete all verbs and adjectives
4. Convert words into their dictionary forms

Why can careless stop-word removal damage sentiment analysis?

1. It can remove negation words that reverse meaning
2. It always increases vocabulary size
3. It converts text into images
4. It prevents the model from calculating accuracy

What is the main difference between stemming and lemmatization?

1. Stemming uses rough word reduction, while lemmatization aims for a valid dictionary base form
2. Stemming works only on images, while lemmatization works on text
3. Lemmatization removes every repeated word
4. There is no difference between them

Which example best demonstrates lemmatization?

1. Reducing “studies” to “studi”
2. Converting “better” to “good” when context supports that lemma
3. Removing “the” from a sentence
4. Splitting a paragraph into sentences

What does TF-IDF attempt to measure?

1. The importance of a term in a document relative to the document collection
2. The emotional tone of an image
3. The duration of an audio recording
4. The number of labels assigned to one observation

In the expression TF-IDF, what does IDF reduce?

1. The influence of terms that appear in many documents
2. The number of documents in the dataset
3. The length of every word
4. The number of output classes

What is a major limitation of a basic Bag-of-Words representation?

1. It cannot represent numerical values
2. It largely ignores word order and contextual meaning
3. It can be used only with audio
4. It automatically leaks the target

What is an embedding?

1. A dense numerical vector representing useful properties or meaning
2. A compressed archive containing the source files
3. A fixed decision threshold
4. A rule for deleting duplicate rows

What distinguishes contextual text embeddings from traditional static word vectors?

1. They can assign different representations to the same word in different contexts
2. They always use only one numerical dimension
3. They remove the need for tokenization
4. They work only for database tables

Why is transfer learning valuable for image classification with a limited labelled dataset?

1. A pretrained model already contains generally useful visual representations
2. It guarantees perfect accuracy
3. It eliminates the need for any labelled examples
4. It converts images into SQL tables automatically

What does a spectrogram represent?

1. How frequency content changes over time in an audio signal
2. The grammatical structure of a sentence
3. The color distribution of an image
4. The class labels of a tabular dataset

What is sentiment analysis?

1. Classifying or scoring the expressed attitude or emotional orientation in content
2. Counting the number of sentences in a document
3. Detecting the file format of an image
4. Grouping transactions by value

A text classifier achieves excellent offline performance because the resolution category was pasted into each ticket before the training export. What happened?

1. The model discovered a useful linguistic pattern
2. The target leaked into the input content
3. The model underfitted
4. The vocabulary was too small

A topic classifier relies heavily on publisher-specific footer text. Why is this dangerous?

1. The footer always contains too many characters
2. The model may be learning a source shortcut instead of genuine topic content
3. Footers cannot be tokenized
4. Topic classification cannot use words

A rare “security incident” class represents 1% of support requests and is almost never detected. What should be reviewed?

1. Overall accuracy only
2. Training accuracy only
3. Per-class recall, confusion matrix, weighting, resampling, and threshold policy
4. The majority classes only

What should happen before classifying scanned document images by their textual content?

1. Feed the compressed document bytes directly into a tabular classifier
2. Apply OCR and inspect the extraction quality
3. Remove all page images without inspection
4. Convert every document manually into a database schema

A CV classifier trained on historical hiring decisions learns to downweight documents mentioning certain universities and activities. What is the most responsible interpretation?

1. The model is automatically fair because it learned from data
2. Removing one explicit protected field guarantees fairness
3. The model may reproduce historical bias and use text as a proxy for sensitive information
4. Adding more records from the same historical process automatically removes the bias

Explain the difference between structured, semi-structured, and unstructured data.

A complete answer should discuss fixed tabular schemas, self-describing tags or keys, and raw content without a rigid field structure.

Explain why unstructured content must be converted into numbers.

A complete answer should mention mathematical model input and provide examples for text, image, and audio representations.

Explain the difference between stemming and lemmatization.

A complete answer should distinguish mechanical word reduction from linguistically informed dictionary-form normalization.

Explain how TF-IDF differs from simple word counts.

A complete answer should explain how TF-IDF reduces the weight of terms that appear across many documents.

Explain the difference between static and contextual embeddings.

A complete answer should explain whether one word receives one fixed representation or a context-dependent representation.

Explain how leakage can occur inside free-form text.

A complete answer should describe post-outcome notes, resolution labels, agent annotations, or other information unavailable when the actual prediction is required.

Explain shortcut learning using an original example.

A complete answer should describe a non-causal artifact that correlates with the class in training data but may disappear in production.

Explain why overall accuracy may be inadequate for rare classes.

A complete answer should mention majority-class dominance, per-class recall, the confusion matrix, and error consequences.

Define an Unstructured Classification Problem

Identify the content type, observation, target, classes, prediction point, intended user, and action supported by the prediction.

Inspect Raw Content

Review representative examples, missing files, corrupted content, language variation, class balance, duplicated content, and labeling consistency.

Create a Text Baseline

Build a leakage-safe pipeline using TF-IDF and a suitable linear classifier. Record class-specific metrics and important errors.

Review Preprocessing Value

Compare the baseline with and without stop-word removal, stemming, or lemmatization. Keep a step only when validation demonstrates a reliable benefit.

Inspect Model Features

Review highly weighted words or tokens and identify leakage, boilerplate, source identifiers, personally identifying information, or unreliable shortcuts.

Design Representative Validation

Decide whether documents should be split randomly, by author, customer, publisher, source, device, speaker, or time period.

Analyze Rare-Class Performance

Report the confusion matrix, per-class precision and recall, macro F1 score, and examples of important false negatives.

Document Responsible-Use Controls

Document data permission, privacy protection, retention, fairness review, human oversight, error handling, monitoring, and model-retirement conditions.

Record Incorrect Answers

List each incorrect question, your selected option, and the correct answer.

Identify the Competency

Use the competency mapping table to connect each mistake to a lesson topic.

Review the Related Lesson

Review the definition, workflow, examples, limitations, and common mistakes.

Explain the Concept Without Notes

Write a short explanation using your own words and a new example.

Apply the Concept

Implement or analyze the concept using a small authorized, public, or synthetic dataset.

Retake the Quiz

Attempt the assessment again without using the answer key.