Sentiment Analysis

Sentiment analysis, also called opinion mining, is the natural language processing task that identifies and classifies the emotional charge or opinion a text expresses, usually as positive, negative or neutral. It serves to know, at scale, how people feel about a product, a brand or a topic.

Levels of detail

It can be analyzed at different granularities. At the document level, a global polarity is assigned to the whole text; at the sentence level, to each sentence. The finest level is the aspect-based one, which distinguishes the opinion about each feature: in “the battery is good but the screen is bad,” it detects a positive sentiment toward the battery and a negative one toward the screen.

How it has evolved

The first systems were lexicon-based: dictionaries that assign a polarity to each word. Then came classic machine learning—algorithms such as Naive Bayes or support vector machines over TF-IDF-type representations. The leap in quality came from deep learning, with recurrent networks, and since 2018 transformers such as BERT, which capture context much better.

Its challenges

It is not a solved problem. Irony and sarcasm invert the literal meaning of words (“great, three hours waiting”). Negation changes the polarity of a whole sentence. Sentiment depends on context and domain: the same word can be praise or criticism depending on where it appears. And there are texts with mixed emotions that resist any single label.

What it is for

Its applications are very commercial: brand monitoring and reputation, the analysis of product reviews, market research, customer service and social media analysis. In finance it is even used to measure market sentiment from news and posts, as one more signal for decision-making. Documentation: academic sentiment-analysis study.