An annotation is additional information connected to a specific part of a document or piece of information. It can be a note that has a comment or explanation. Sometimes, annotations are written in the margins of book pages. For annotations related to different types of digital media, such as websites and text documents, see web annotation and text annotation.

Literature, grammar and educational purposes

Annotation practices involve highlighting a phrase or sentence and adding a comment, circling a word that needs explaining, asking a question when something is unclear, or writing a short summary of an important part. These practices also encourage students to create a history by using materials and doing hands-on annotation activities.

Text and film annotation is a method where people add comments or notes inside a film or text. When analyzing videos, researchers must consider their own ideas and beliefs first. Annotations can be added during the video recording process and placed within the video itself. They help people write down their thoughts and feelings. This method can be used at any stage of analysis and can include more notes later. Anthropologist Clifford Geertz called this process a "thick description," which shows how helpful annotations can be, especially when explaining how they are used in films.

Marginalia are notes or drawings written in the margins of books. Readers often added these notes to help future readers understand the text better.

Textual scholarship is a field that studies texts and adds historical information to help people understand them more clearly.

Students often highlight important parts of books to interact with the text. Annotations can help them find key phrases quickly or add notes in the margins to study better and connect ideas in the text to their own knowledge or recurring themes. Annotation tasks are often given to high school students to help them focus on the material.

Annotated bibliographies include comments about how useful or good each source is, in addition to basic information like the title and author.

Students use annotations for schoolwork and to understand their own thoughts, feelings, and emotions.

Mathematical symbols and formulas can be explained with words to avoid confusion, since symbols might mean different things (for example, "E" can mean "energy" or "expectation value"). Tools like "AnnoMathTeX" can help make this process faster.

From a learning perspective, annotations help students pay attention to important parts of pictures or diagrams. Experts, like doctors, use annotations to explain their interpretations of images to others. This helps people with different knowledge levels understand each other better. Studies have shown that using annotations with images and speech during online meetings improves communication.

On January 15, 2019, YouTube removed annotations after about 10 years of use. Annotations allowed users to add notes that appeared during videos, but YouTube said they didn’t work well on small screens and were sometimes misused.

Software and engineering

Markup languages like XML and HTML add notes to text in a way that is clearly different from the text itself. These languages can be used to describe how text should look visually or to provide information that computers can understand, such as in the semantic web.

This includes formats like CSV and XLS. The process of adding labels that explain the meaning of data in tables is called semantic labeling. Semantic labeling involves using information from organized systems of knowledge to assign labels to table data. This process is also called semantic annotation. Semantic labeling is often done automatically or with some help from people. Techniques for semantic labeling work with columns that contain names, numbers, coordinates, and other types of data.

There are several types of semantic labeling that use machine learning. These methods can be grouped based on the work of Flach as follows: geometric (using lines and shapes, such as Support-vector machine and Linear regression), probabilistic (such as Conditional random field), logical (such as Decision tree learning), and Non-ML techniques (such as balancing coverage and specificity). Note that geometric, probabilistic, and logical machine learning models can overlap.

Pham et al. use the Jaccard index and TF-IDF similarity for text data and the Kolmogorov–Smirnov test for numeric data. Alobaid and Corcho use fuzzy clustering (c-means) to label numeric columns.

Limaye et al. use TF-IDF similarity and graphical models. They also use support-vector machine to calculate weights. Venetis et al. create an isA database that includes pairs of (instance, class) and use these pairs to calculate maximum likelihood. Alobaid and Corcho approximate the q-q plot to predict the properties of numeric columns.

Syed et al. built Wikitology, which is "a hybrid knowledge base of structured and unstructured information from Wikipedia enhanced with RDF data from DBpedia and other linked data sources." For the Wikitology index, they use PageRank for Entity linking, a task often used in semantic labeling. Since they could not query Google for all Wikipedia articles to get PageRank, they used Decision tree to approximate it.

Alobaid and Corcho presented a method to label columns with names. The process begins by matching names in the column to entities in a reference knowledge graph (e.g., DBpedia). Classes are then collected, and each class is scored based on formulas that consider how often each class appears and its position in the subClass hierarchy.

Common semantic labeling tasks include:

This is the most common task in semantic labeling. Given a text from a cell and a data source, the method predicts the entity and links it to the one found in the data source. For example, if the input were the text "Richard Feynman" and a URL to the SPARQL endpoint of DBpedia, the method would return "http://dbpedia.org/resource/Richard_Feynman," which is the entity from DBpedia. Some methods use exact matches, while others use similarity metrics like Cosine similarity.

The subject column of a table contains the main subjects or entities in the table. Some methods require the subject column as input, while others predict it, such as TableMiner+.

Columns are divided differently by various methods. Some group them into text and numbers, while others further divide them (e.g., Number Typology, Date, coordinates).

The relationship between Madrid and Spain is "capitalOf." Such relationships are often found in ontologies like DBpedia. Venetis et al. use TextRunner to extract relationships between columns. Syed et al. use the relationship between entities in two columns and select the most frequent one.

T2D is the most common standard for evaluating semantic labeling. Two versions of T2D exist: T2Dv1 (sometimes called T2D) and T2Dv2. Other known benchmarks include those from the SemTab Challenge.

The "annotate" function (also called "blame" or "praise") in source control systems like Git, Team Foundation Server, and Subversion identifies who made changes to the source code. This function creates a copy of the source code where each line is labeled with the name of the last person who edited it (and possibly a revision number). This helps track who made changes if a problem occurs or identify the author of useful code.

In Java, annotations are a type of metadata added to code and can be used with reflective programming. Annotations can be added to classes, methods, variables, parameters, and packages. They can be stored in class files and used by the Java virtual machine, which can affect how an application runs at runtime. Java allows creating new annotations based on existing ones.

Other languages, like C#, have similar features called "attributes." C++ also has "attributes" that give instructions to the compiler, and C++26 introduces reflection annotations similar to Java annotations.

Automatic image annotation is used to classify images for image retrieval systems.

Since the 1980s, molecular biology and bioinformatics have needed DNA annotation. DNA annotation, or genome annotation, is the process of identifying the locations of genes and their functions in a genome. An annotation is a note added to explain or comment on something. Once a genome is sequenced, it must be annotated to understand its contents.

In the digital imaging field, "annotation" often refers to visible metadata added to an image without altering the original image, such as sticky notes, virtual laser pointers, circles, arrows, and black-outs (similar to redaction).

In the medical imaging field, an annotation is often called a region of interest and is encoded in DICOM format.

Other uses

In the United States, legal publishers such as Thomson West and LexisNexis create versions of laws with added information about court cases that explain the laws. Both the national United States Code and state laws can be explained by courts, and these annotated versions are helpful for legal research.

One goal of adding information is to prepare data for use with computers. Before adding information, a plan is made that usually includes labels. When adding labels, workers manually place them in text where specific language details are found using an editor. The plan makes sure labels are used the same way throughout the data and allows checking of previous labels. In addition to labels, more detailed forms of adding information include labeling phrases and relationships, such as in treebanks. Many different ways of adding information have been created, along with various formats and tools for managing these annotations, as explained in the Linguistic Annotation Wiki.