Sentiment Analysis

Sentiment analysis scores the opinion or emotional tone in text, usually as positive, negative, or neutral. Stronger systems also detect specific emotions, aspect-level views, or intensity.

Document-level classification asks whether a whole review is positive or negative. Aspect-based analysis is finer. "The food was excellent but the service was slow" is positive about food and negative about service. Emotion detection names joy, anger, fear, sadness, surprise, or disgust rather than polarity alone. Training data often comes from labeled reviews, social posts, or surveys.

Transfer learning from large language models raised accuracy on subtle or indirect sentiment.

Sarcasm uses positive words with negative meaning. Implicit sentiment hides in facts that imply an opinion. Words change connotation by domain. Culture shifts the reading. Uses include brand monitoring, customer feedback, social analytics, investor-sentiment gauges in markets, political analysis, and support queues. Scores can be rolled up over time, topics, or demographics to watch trends.

Polarity is the coarse task: positive, negative, neutral. Aspect-based systems attach those labels to targets such as food versus service in one review. Emotion taxonomies add joy, anger, fear, sadness, surprise, and disgust. Fine-tuned language models handle indirect sentiment better than bag-of-words, but sarcasm and domain jargon still break them.

A word that is praise in restaurants can be a complaint in aviation. Teams aggregate scores by week, product, or region for brand and market monitoring, including investor-sentiment dashboards. Sentiment models score whether text is positive or negative. The Stanford Sentiment Treebank is a standard dataset.