Multi-class text classification in NLP - Model Metrics & Evaluation

Metrics & Evaluation - Multi-class text classification

Which metric matters for Multi-class text classification and WHY

In multi-class text classification, the goal is to correctly assign each text to one category out of many. The key metrics are accuracy, precision, recall, and F1-score calculated per class and averaged. Accuracy shows overall correct predictions. Precision tells us how many predicted texts for a class truly belong there. Recall shows how many texts of a class were found by the model. F1-score balances precision and recall. These metrics help us understand if the model is good at finding and correctly labeling each class.

Confusion matrix for Multi-class classification

Here is an example confusion matrix for 3 classes: Sports, Politics, and Tech.

          Predicted
          S   P   T
    True S 50  2   3
         P  4 45   1
         T  2  3  48
    

Explanation:

• 50 texts truly Sports were predicted as Sports (True Positives for Sports)
• 2 Sports texts were wrongly predicted as Politics (False Negatives for Sports)
• 3 Sports texts were wrongly predicted as Tech (False Negatives for Sports)
• Similarly for Politics and Tech classes.

From this matrix, we calculate precision, recall, and F1 for each class.

Precision vs Recall tradeoff with examples

Imagine a model classifying news articles into categories. If the model has high precision for Sports, it means when it says an article is Sports, it is usually right. But if recall is low, it misses many Sports articles.

For example, if you want to recommend Sports articles only when very sure, prioritize precision. If you want to catch all Sports articles even if some mistakes happen, prioritize recall.

Balancing precision and recall with F1-score helps when both false positives and false negatives matter.

What good vs bad metric values look like

Good metrics:

• Accuracy above 80% on balanced data
• Precision and recall above 75% for each class
• F1-score close to precision and recall, showing balance

Bad metrics:

• Accuracy near random guess (e.g., 33% for 3 classes)
• Very low recall for some classes (missing many texts)
• High precision but very low recall or vice versa (unbalanced)

Common pitfalls in metrics for multi-class text classification

• Accuracy paradox: High accuracy can be misleading if classes are imbalanced. For example, if one class is 90% of data, predicting it always gives 90% accuracy but poor performance on others.
• Ignoring per-class metrics: Overall accuracy hides poor results on minority classes.
• Data leakage: If test data leaks into training, metrics look unrealistically high.
• Overfitting: Very high training accuracy but low test accuracy means model memorizes training data, not generalizing.

Self-check question

Your multi-class text classifier has 98% accuracy but recall for one important class is only 12%. Is this model good for production? Why or why not?

Answer: No, it is not good. The high accuracy likely comes from correctly predicting the majority classes. But the very low recall for the important class means the model misses most texts of that class. This can cause serious problems if that class is critical. You should improve recall for that class before using the model.

Key Result

In multi-class text classification, balanced precision, recall, and F1-score per class are key to understanding model quality beyond overall accuracy.

Practice

1. What is the main goal of multi-class text classification?

easy

A. To sort text into multiple categories based on content

B. To translate text into another language

C. To count the number of words in a text

D. To generate new text from a given input

Solution

1. Step 1: Understand the task of multi-class text classification

   This task involves assigning each text sample to one category out of many possible categories.

2. Step 2: Compare options with the task definition

   Only To sort text into multiple categories based on content describes sorting text into multiple categories, which matches the task.

3. Final Answer:

   To sort text into multiple categories based on content -> Option A

4. Quick Check:

   Multi-class classification = sorting into many categories [OK]

Hint: Multi-class means sorting text into many groups [OK]

Common Mistakes:

• Confusing classification with translation
• Thinking it counts words instead of categorizing
• Mixing generation with classification

2. Which of the following is the correct way to represent text data for multi-class classification?

easy

A. Converting text into numerical vectors like TF-IDF or embeddings

B. Using raw text strings directly as input to the model

C. Sorting text alphabetically before training

D. Removing all punctuation and spaces only

Solution

1. Step 1: Identify how models process text

   Models cannot understand raw text strings; they need numbers to learn patterns.

2. Step 2: Check which option converts text to numbers

   Converting text into numerical vectors like TF-IDF or embeddings mentions converting text into numerical vectors like TF-IDF or embeddings, which is correct.

3. Final Answer:

   Converting text into numerical vectors like TF-IDF or embeddings -> Option A

4. Quick Check:

   Text must be numbers for models [OK]

Hint: Models need numbers, not raw text, to learn [OK]

Common Mistakes:

• Feeding raw text directly to models
• Thinking sorting text helps classification
• Ignoring the need for numerical representation

3. Given the following Python code snippet for multi-class text classification, what will be the output of print(predicted_class)?

from sklearn.feature_extraction.text import CountVectorizer
from sklearn.naive_bayes import MultinomialNB

texts = ["I love cats", "Dogs are great", "I hate rain"]
labels = ["positive", "positive", "negative"]

vectorizer = CountVectorizer()
X = vectorizer.fit_transform(texts)

model = MultinomialNB()
model.fit(X, labels)

new_text = ["I love dogs"]
X_new = vectorizer.transform(new_text)
predicted_class = model.predict(X_new)[0]

medium

A. "neutral"

B. "negative"

C. An error because of unknown words

D. "positive"

Solution

1. Step 1: Understand training data and labels

   The model is trained on texts labeled as "positive" or "negative". "I love cats" and "Dogs are great" are positive, "I hate rain" is negative.

2. Step 2: Predict class for new text "I love dogs"

   The new text contains words "I", "love", and "dogs" which appear in positive examples. The model predicts "positive" as the class.

3. Final Answer:

   "positive" -> Option D

4. Quick Check:

   New text matches positive words, so prediction is positive [OK]

Hint: New text similar to positive examples predicts positive [OK]

Common Mistakes:

• Assuming unknown words cause errors
• Choosing negative because of 'dogs' only
• Picking neutral which is not a trained label

4. Identify the error in this multi-class text classification code snippet:

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression

texts = ["happy day", "sad night", "joyful morning"]
labels = ["positive", "negative", "positive"]

vectorizer = TfidfVectorizer()
X = vectorizer.fit_transform(texts)

model = LogisticRegression()
model.fit(texts, labels)

medium

A. Labels should be numbers, not strings

B. Passing raw texts instead of vectorized data to model.fit

C. Using LogisticRegression instead of Naive Bayes

D. TfidfVectorizer cannot be used with LogisticRegression

Solution

1. Step 1: Check input to model.fit()

   The model.fit() method expects numerical features, but raw texts are passed instead of vectorized data.

2. Step 2: Identify correct input

   The vectorized data X should be passed to model.fit, not the original texts.

3. Final Answer:

   Passing raw texts instead of vectorized data to model.fit -> Option B

4. Quick Check:

   Model needs numbers, not raw text, for training [OK]

Hint: Model.fit needs vectorized data, not raw text [OK]

Common Mistakes:

• Passing raw text directly to model.fit
• Thinking label type causes error here
• Believing vectorizer choice causes this error

5. You have a dataset with 5 classes and highly imbalanced text samples per class. Which approach best improves multi-class classification performance?

hard

A. Use only the most frequent class for prediction

B. Ignore imbalance and train model on raw data

C. Use class weighting or oversampling to balance training data

D. Remove classes with fewer samples to simplify problem

Solution

1. Step 1: Understand class imbalance impact

   Imbalanced classes cause models to favor majority classes, reducing accuracy on minority classes.

2. Step 2: Identify best practice to handle imbalance

   Using class weighting or oversampling balances the training data, helping the model learn all classes better.

3. Final Answer:

   Use class weighting or oversampling to balance training data -> Option C

4. Quick Check:

   Balance data to improve multi-class model accuracy [OK]

Hint: Balance classes with weighting or oversampling [OK]

Common Mistakes:

• Ignoring imbalance and expecting good results
• Removing minority classes loses valuable data
• Predicting only the majority class ignores others