Monitoring NLP models - ML Experiment: Train & Evaluate

Experiment - Monitoring NLP models

Problem:You have a text classification NLP model deployed to classify customer reviews as positive or negative. The model was trained well, but after deployment, its performance might degrade over time due to changes in language or topics.

Current Metrics:Training accuracy: 92%, Validation accuracy: 88%, Current deployed model accuracy on recent data: 75%

Issue:The model shows signs of performance degradation (accuracy dropped from 88% validation to 75% on recent data), indicating possible data drift or concept drift.

Your Task

Implement a monitoring system that tracks the model's prediction accuracy on new incoming data and alerts when accuracy drops below 85%.

You cannot retrain the model in this task.

Use only Python and common NLP libraries (e.g., scikit-learn, pandas).

Simulate new incoming data with a small sample.

Hint 1

Hint 2

Hint 3

Solution

NLP

import pandas as pd
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score

# Simulated training data
train_texts = ["I love this product", "This is bad", "Amazing quality", "Not good", "Excellent!", "Terrible experience"]
train_labels = [1, 0, 1, 0, 1, 0]

# Train a simple model
vectorizer = CountVectorizer()
X_train = vectorizer.fit_transform(train_texts)
model = LogisticRegression(max_iter=1000)
model.fit(X_train, train_labels)

# Simulated new incoming data batch (recent data)
new_texts = ["I hate this", "Very good", "Worst ever", "I like it", "Not bad"]
new_labels = [0, 1, 0, 1, 1]  # True labels for monitoring

# Transform new data
X_new = vectorizer.transform(new_texts)

# Predict with deployed model
predictions = model.predict(X_new)

# Compute accuracy
accuracy = accuracy_score(new_labels, predictions) * 100

# Monitoring check
threshold = 85.0
print(f"Model accuracy on new data: {accuracy:.2f}%")
if accuracy < threshold:
    print(f"ALERT: Model accuracy dropped below {threshold}%!")

Added a monitoring function to compute accuracy on new incoming data.

Simulated new data batch with true labels to evaluate model performance.

Included an alert print statement when accuracy falls below 85%.

Results Interpretation

Before monitoring: Model accuracy on recent data was unknown or not tracked, leading to unnoticed performance drop.

After monitoring: Model accuracy on new data is computed as 80%, which is below the 85% threshold, triggering an alert.

Monitoring deployed NLP models with simple accuracy checks on new data helps detect performance drops early, enabling timely interventions before serious issues occur.

Bonus Experiment

Extend the monitoring system to track and plot accuracy over multiple batches of new data to visualize trends.

💡 Hint

Store accuracy values in a list and use matplotlib to plot accuracy over time.

Practice

1. Why is monitoring important for NLP models in production?

easy

A. To ensure the model stays accurate and reliable over time

B. To make the model run faster on the user's device

C. To reduce the size of the model file

D. To increase the number of features in the model

Solution

1. Step 1: Understand the purpose of monitoring

   Monitoring tracks model performance to detect when it degrades or behaves unexpectedly.

2. Step 2: Relate monitoring to model reliability

   Keeping the model accurate and reliable ensures users get correct results consistently.

3. Final Answer:

   To ensure the model stays accurate and reliable over time -> Option A

4. Quick Check:

   Monitoring = Accuracy and reliability [OK]

Hint: Monitoring checks if model predictions stay correct over time [OK]

Common Mistakes:

• Confusing monitoring with model training
• Thinking monitoring changes model size
• Believing monitoring speeds up the model

2. Which metric is commonly used to monitor the accuracy of an NLP classification model?

easy

A. Latency

B. Recall

C. Model size

D. Training time

Solution

1. Step 1: Identify metrics related to classification quality

   Recall measures how many relevant items the model correctly finds, important for classification.

2. Step 2: Differentiate from other metrics

   Latency measures speed, model size and training time are unrelated to accuracy.

3. Final Answer:

   Recall -> Option B

4. Quick Check:

   Recall = Accuracy metric [OK]

Hint: Recall measures correct positive predictions [OK]

Common Mistakes:

• Choosing latency as accuracy metric
• Confusing model size with performance
• Selecting training time instead of recall

3. Given this monitoring alert rule:
if accuracy < 0.85 then alert('Low accuracy')
What happens if the model accuracy drops to 0.80?

medium

A. No alert is triggered

B. The system shuts down

C. The model automatically retrains

D. An alert 'Low accuracy' is triggered

Solution

1. Step 1: Understand the alert condition

   The alert triggers when accuracy is less than 0.85.

2. Step 2: Check the given accuracy value

   Accuracy is 0.80, which is less than 0.85, so the condition is true.

3. Final Answer:

   An alert 'Low accuracy' is triggered -> Option D

4. Quick Check:

   Accuracy 0.80 < 0.85 triggers alert [OK]

Hint: Alert triggers when metric is below threshold [OK]

Common Mistakes:

• Thinking alert triggers only if accuracy equals 0.85
• Assuming model retrains automatically
• Believing system shuts down on alert

4. You set up a latency alert for your NLP model:
if latency > 200ms then alert('High latency')
But no alert triggers even when users report slow responses. What is the likely problem?

medium

A. The latency threshold is set too high

B. The alert message text is incorrect

C. Latency is measured in seconds, not milliseconds

D. The model accuracy is too low

Solution

1. Step 1: Analyze the alert condition and user reports

   The alert triggers if latency is above 200ms, but users report slow responses.

2. Step 2: Consider threshold setting

   If users feel slow but latency is below 200ms, threshold is too high to catch issues.

3. Final Answer:

   The latency threshold is set too high -> Option A

4. Quick Check:

   High threshold misses slow responses [OK]

Hint: Check if alert thresholds match user experience [OK]

Common Mistakes:

• Changing alert text without fixing threshold
• Confusing latency units
• Blaming accuracy for latency issues

5. You want to monitor an NLP model's performance over time and detect sudden drops in accuracy. Which approach is best?

hard

A. Retrain the model daily without monitoring

B. Only monitor latency since accuracy is stable

C. Set a fixed accuracy threshold and alert when accuracy falls below it

D. Ignore monitoring and rely on user feedback

Solution

1. Step 1: Identify the goal of monitoring

   The goal is to detect sudden drops in accuracy to maintain model quality.

2. Step 2: Evaluate each option

   Setting a fixed threshold and alerting is a proactive way to catch drops. Other options ignore monitoring or focus on unrelated metrics.

3. Final Answer:

   Set a fixed accuracy threshold and alert when accuracy falls below it -> Option C

4. Quick Check:

   Threshold alerts catch accuracy drops [OK]

Hint: Use thresholds to catch sudden accuracy drops early [OK]

Common Mistakes:

• Ignoring accuracy monitoring
• Relying only on latency
• Skipping alerts and waiting for user reports