How Sentiment Analysis Works

Collect and Label Training Data

Every supervised sentiment analysis system starts with labeled training data: text samples that humans have tagged with their sentiment. The most common sources are product reviews from sites like Amazon or Yelp, where the star rating provides a natural sentiment label (1-2 stars = negative, 4-5 stars = positive, 3 stars = neutral or excluded). Movie reviews from IMDB's dataset of 50,000 labeled reviews have been a standard benchmark since 2011. Twitter datasets with hashtag-based sentiment labels (#happy, #angry, #sad) provide social media training data, though they tend to be noisier.

The quality of training data determines the ceiling of model performance. Labels must be consistent: if some annotators consider "it was okay" positive and others consider it neutral, the model receives conflicting signals. Inter-annotator agreement rates typically range from 80% to 90% for binary sentiment (positive/negative), reflecting genuine ambiguity in human judgment. For fine-grained sentiment (very positive, positive, neutral, negative, very negative), agreement rates drop to 60% to 70%. This human disagreement sets an upper bound on what any model can achieve, because even perfect models cannot be more consistent than the humans who provided the labels.

Domain matters enormously. A sentiment model trained on movie reviews performs poorly on financial news because the same words carry different connotations. "Volatile" is neutral in a movie review but negative in financial commentary. "Aggressive" might be positive in a sports context ("aggressive strategy") but negative in a customer service context. Effective sentiment analysis requires either training data from the target domain or a domain adaptation strategy that fine-tunes a general model on a smaller domain-specific dataset.

Preprocess Text for Analysis

Raw text must be cleaned and standardized before a model can learn from it. Tokenization splits the text into words or subword units. Lowercasing normalizes "Great," "great," and "GREAT" into a single form, though all-caps may carry sentiment information (shouting) that lowercasing discards. Stopword removal eliminates common words like "the," "is," and "and" that carry little sentiment signal. Stemming or lemmatization reduces words to their base forms, so "running," "ran," and "runs" all map to "run."

Negation handling is critical and surprisingly tricky. The sentence "This movie is not good" has the opposite sentiment of "This movie is good," but a bag-of-words model sees both sentences as containing "good" and might classify them identically. Simple negation handling attaches a "NOT_" prefix to all words between a negation word and the next punctuation mark, so "not good at all" becomes "not NOT_good NOT_at NOT_all." More sophisticated approaches use syntactic parsing to determine the scope of negation, handling cases like "I don't think this is a bad idea" where double negation creates a positive sentiment.

For modern transformer-based models, minimal preprocessing is needed because the model learns to handle these variations during pre-training. The tokenizer handles subword splitting, and the model's attention mechanism captures negation, intensification, and other contextual modifiers without explicit preprocessing rules. However, tasks like emoji handling, URL removal, and user mention anonymization in social media text remain important preprocessing steps even for transformer models.

Train a Sentiment Classifier

The modern approach to sentiment classification is fine-tuning a pre-trained language model. Starting with a model like BERT, RoBERTa, or a similar transformer that has been pre-trained on billions of words, you add a classification layer on top and train the entire system on your labeled sentiment data. The pre-trained weights already encode deep understanding of language structure and word meaning, so fine-tuning requires relatively little labeled data (a few thousand examples often suffice) and converges quickly (2 to 4 epochs of training).

The fine-tuning process works by feeding labeled examples through the model, computing the cross-entropy loss between the predicted sentiment probabilities and the true labels, and using backpropagation to adjust all the model's weights. A learning rate of 2e-5 to 5e-5 is standard for fine-tuning transformers, much smaller than the learning rates used in pre-training, to avoid catastrophically overwriting the pre-trained knowledge. Training typically takes 30 minutes to 2 hours on a single GPU for datasets of 10,000 to 100,000 examples.

Evaluation uses held-out test data that the model has never seen during training. Standard metrics include accuracy (percentage of correctly classified examples), precision (of the examples the model labeled positive, what fraction actually are positive), recall (of the actually positive examples, what fraction did the model catch), and F1 score (the harmonic mean of precision and recall). On the SST-2 binary sentiment benchmark, state-of-the-art models achieve over 96% accuracy. On more challenging multi-class or fine-grained benchmarks, accuracy ranges from 55% to 75%, depending on the granularity of the sentiment scale and the difficulty of the dataset.

Handle Edge Cases

Sarcasm is the hardest edge case for sentiment analysis. "Oh great, another meeting" expresses the opposite of what "great" literally means. Detecting sarcasm requires understanding that the speaker's true attitude contradicts their words, which in turn requires understanding social context, speaker expectations, and cultural norms. Current models detect obvious sarcasm in some cases but miss subtle instances regularly. Sarcasm detection accuracy on benchmark datasets hovers around 75% to 80%, far below overall sentiment accuracy.

Mixed sentiment occurs when a text contains both positive and negative opinions. "The food was excellent but the service was terrible" is simultaneously positive about food and negative about service. Document-level sentiment analysis forces this into a single label, losing important information. Aspect-based sentiment analysis (ABSA) solves this by identifying the aspects mentioned in the text (food, service) and assigning sentiment to each one independently. This provides much more useful information for businesses: knowing that food is praised while service is criticized is far more actionable than knowing the review is "mixed."

Comparative sentiment presents another challenge. "iPhone has a better camera than Samsung" is positive about iPhone and negative about Samsung, but a simple classifier might label it positive overall without distinguishing the targets. Implicit sentiment, where the opinion is conveyed through description rather than evaluative language, is also difficult: "I waited three hours for my food" expresses clear negative sentiment without using any traditionally negative words.

Deploy and Monitor

Production sentiment analysis systems must handle high throughput with low latency. A social media monitoring tool might process millions of posts per day. A customer support system needs real-time sentiment scoring to route urgent complaints to human agents. Model distillation, where a large, accurate model is used to train a smaller, faster model, is commonly used to balance accuracy and speed. Quantized models that use 8-bit instead of 32-bit weights reduce memory requirements and increase throughput with minimal accuracy loss.

Monitoring model performance over time is essential because language evolves. New slang, emerging cultural references, and shifts in how people express opinions can cause model accuracy to degrade. The word "sick" meant negative a generation ago but is now commonly used positively in informal language. "Mid" emerged as a negative sentiment indicator that older models would not recognize. Periodic re-evaluation on fresh annotated data and retraining on recent text keeps the model aligned with current language use.