Domain-specific sentiment in NLP - ML Experiment: Train & Evaluate

Experiment - Domain-specific sentiment

Problem:You want to build a sentiment analysis model that understands the feelings expressed in movie reviews. The current model is trained on general product reviews and does not perform well on movie reviews.

Current Metrics:Training accuracy: 92%, Validation accuracy: 68%, Validation loss: 0.85

Issue:The model is overfitting to general product reviews and does not generalize well to movie reviews, resulting in low validation accuracy.

Your Task

Reduce overfitting and improve validation accuracy on movie reviews to at least 80%, while keeping training accuracy below 90%.

You can only modify the model architecture and training process.

You cannot change the dataset or add more data.

You must keep the training time reasonable (under 10 minutes).

Hint 1

Hint 2

Hint 3

Hint 4

Solution

NLP

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Embedding, LSTM, Dense, Dropout
from tensorflow.keras.callbacks import EarlyStopping

# Assume X_train, y_train, X_val, y_val are preprocessed movie review data

model = Sequential([
    Embedding(input_dim=10000, output_dim=128, input_length=100),
    LSTM(64, return_sequences=False),
    Dropout(0.5),
    Dense(32, activation='relu'),
    Dropout(0.5),
    Dense(1, activation='sigmoid')
])

model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
              loss='binary_crossentropy',
              metrics=['accuracy'])

early_stop = EarlyStopping(monitor='val_loss', patience=3, restore_best_weights=True)

history = model.fit(X_train, y_train,
                    epochs=20,
                    batch_size=32,
                    validation_data=(X_val, y_val),
                    callbacks=[early_stop])

Added Dropout layers after LSTM and Dense layers to reduce overfitting.

Used EarlyStopping callback to stop training when validation loss stops improving.

Set learning rate to 0.001 for better convergence.

Kept model size moderate to avoid overfitting.

Results Interpretation

Before: Training accuracy: 92%, Validation accuracy: 68%, Validation loss: 0.85

After: Training accuracy: 88%, Validation accuracy: 82%, Validation loss: 0.55

Adding dropout and early stopping helped reduce overfitting, improving validation accuracy on domain-specific movie reviews while keeping training accuracy reasonable.

Bonus Experiment

Try using pre-trained word embeddings like GloVe or Word2Vec fine-tuned on movie reviews to further improve validation accuracy.

💡 Hint

Load pre-trained embeddings and set them as weights in the Embedding layer with trainable=True to adapt to movie review language.

Practice

1. What is the main advantage of using domain-specific sentiment analysis over general sentiment analysis?

easy

A. It works for all topics equally well.

B. It requires no training data.

C. It ignores the context of words.

D. It understands feelings better in a specific area.

Solution

1. Step 1: Understand domain-specific sentiment

   Domain-specific sentiment focuses on feelings related to a particular topic or area, making it more precise.

2. Step 2: Compare with general sentiment

   General sentiment tries to work on all topics but may miss nuances in specialized areas.

3. Final Answer:

   It understands feelings better in a specific area. -> Option D

4. Quick Check:

   Domain focus improves understanding = C [OK]

Hint: Domain-specific means better feelings understanding in one area [OK]

Common Mistakes:

• Thinking it needs no training data
• Assuming it works equally well everywhere
• Believing it ignores word context

2. Which of the following is the correct way to prepare data for domain-specific sentiment training?

easy

A. Collect labeled data from the target domain.

B. Train on unlabeled data from a different domain.

C. Use only positive reviews from all domains.

D. Use random text from any topic without labels.

Solution

1. Step 1: Identify training data needs

   Domain-specific sentiment requires labeled examples from the target domain to learn correctly.

2. Step 2: Evaluate options

   Only collecting labeled data from the target domain provides labeled examples from the correct domain, which is essential for training.

3. Final Answer:

   Collect labeled data from the target domain. -> Option A

4. Quick Check:

   Labeled target data needed = D [OK]

Hint: Training needs labeled data from the right domain [OK]

Common Mistakes:

• Using unlabeled or random data
• Mixing data from unrelated domains
• Ignoring the need for labels

3. Given this Python snippet for domain-specific sentiment prediction:

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

texts = ['Great battery life', 'Poor screen quality', 'Excellent camera']
labels = [1, 0, 1]  # 1=positive, 0=negative

vectorizer = CountVectorizer()
X = vectorizer.fit_transform(texts)
model = LogisticRegression()
model.fit(X, labels)

new_text = ['Battery lasts long']
X_new = vectorizer.transform(new_text)
pred = model.predict(X_new)

What is the expected output of pred?

medium

A. [1]

B. [0]

C. Error due to missing labels

D. [1, 0]

Solution

1. Step 1: Understand training data and labels

   The model is trained on positive and negative examples related to product features.

2. Step 2: Predict sentiment for new text

   'Battery lasts long' is similar to 'Great battery life', which is labeled positive (1), so prediction should be positive.

3. Final Answer:

   [1] -> Option A

4. Quick Check:

   Similar positive text predicts 1 = A [OK]

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

Common Mistakes:

• Expecting multiple predictions for single input
• Confusing labels or expecting error
• Ignoring vectorizer transform step

4. You have this code snippet for domain-specific sentiment training:

texts = ['Good food', 'Bad service']
labels = [1, 0]
vectorizer = CountVectorizer()
X = vectorizer.fit_transform(texts)
model = LogisticRegression()
model.fit(X, labels)

new_text = ['Bad food']
X_new = vectorizer.transform(new_text)
pred = model.predict(X_new)
print(pred)

The output is always [1] even for negative phrases. What is the likely error?

medium

A. Labels are reversed in training data.

B. The vectorizer was not fit before transform.

C. The model was trained on too few examples.

D. The new text was not transformed correctly.

Solution

1. Step 1: Check training data size

   Only two examples are used, which is too small for the model to learn properly.

2. Step 2: Analyze model behavior

   With limited data, the model may predict the majority class or fail to distinguish negative phrases.

3. Final Answer:

   The model was trained on too few examples. -> Option C

4. Quick Check:

   Small training data causes poor predictions = A [OK]

Hint: Too few training examples cause wrong predictions [OK]

Common Mistakes:

• Assuming vectorizer not fit causes this
• Thinking labels are reversed
• Believing transform step is incorrect

5. You want to improve domain-specific sentiment analysis for movie reviews. Which approach best combines domain knowledge and model accuracy?

hard

A. Train a sentiment model on general tweets and apply it to movie reviews.

B. Collect labeled movie reviews, fine-tune a pre-trained language model, and test on movie data.

C. Use a dictionary of positive and negative words from unrelated domains.

D. Train a model only on unlabeled movie reviews using clustering.

Solution

1. Step 1: Identify domain-specific data needs

   Using labeled movie reviews ensures the model learns relevant sentiment patterns.

2. Step 2: Use advanced model fine-tuning

   Fine-tuning a pre-trained language model adapts general knowledge to the movie domain, improving accuracy.

3. Final Answer:

   Collect labeled movie reviews, fine-tune a pre-trained language model, and test on movie data. -> Option B

4. Quick Check:

   Labeled domain data + fine-tuning = best accuracy [OK]

Hint: Fine-tune with labeled domain data for best results [OK]

Common Mistakes:

• Using unrelated domain data only
• Relying on unlabeled data without supervision
• Using generic word lists without context