Annotation

Annotation is the process of labelling raw data — text, images, audio, video — so that a machine learning model can learn from it. Without annotation, most AI training would be impossible.

How annotation works

Imagine you want to train a model to detect cats in photos. You need thousands of images where a human has marked which ones contain cats and which do not. That marking process is annotation. For more complex tasks, annotation might involve drawing bounding boxes around objects, highlighting specific words in a sentence, or transcribing spoken words from audio.

Types of annotation

• Text annotation includes labelling sentiment (positive, negative, neutral), tagging named entities (people, places, companies), and marking intent in customer queries
• Image annotation includes bounding boxes, polygon outlines, key-point marking, and pixel-level segmentation
• Audio annotation includes transcription, speaker identification, and emotion labelling
• Video annotation tracks objects frame by frame for tasks like autonomous driving

Who does the annotating

Annotation is often done by large teams of human workers, sometimes called data labellers. Companies like Scale AI and Labelbox provide annotation platforms and workforces. Increasingly, AI-assisted annotation speeds up the process: a model generates initial labels, and humans correct them.

Quality matters enormously

The quality of annotations directly determines the quality of the trained model. Ambiguous labels, inconsistent standards, or careless annotation create noisy training data that leads to unreliable models. This is why annotation guidelines — clear rules for how to label each data point — are critical.

The cost of annotation

Annotation is one of the most expensive and time-consuming parts of building AI. It is often the bottleneck in AI projects, especially for specialised domains like medical imaging where expert annotators are required.