Syntactic annotation includes across-word analysis and its goal is to describe structure of sentences. It shows how the words build phrases and sentences, and their relationship with each other in a sentence.

Sometimes a syntactically annotated corpus is called a treebank, because it is represented in a tree structure.

A distinction is made between two basic models: constituency- and dependency-based structure annotation. Both structures give the hierarchical structure of the sentence. They differ from each other in the way that the elements hierarchically relate to each other in one sentence. In the constituent structure, the words build a syntactic unit such as noun phrase and verb phrase. The phrases are hierarchically arranged and may be deeply nested.

However, the dependence structure is based on the dependency between the words in one sentence and it describes how words relate to other words. In a dependence structure, one word is a head and others are its dependents. Usually the tensed verb is head of structure and the other words are subordinated to it.

The further distinction is made between partial and full syntactic analysis. The full parsing describes a complete parse tree of a sentence. This kind of parsing is difficult and time consuming, because it has large grammatical coverage. Additionally it is obtained with a very large technical and time effort. The reason is that all possible parse trees must be calculated for a given sentence.

Another alternative analysis is the partial parsing or sometimes called chunking. It describes a partial structure of a sentence without building a complete parse tree. In contrast to full syntactic analysis, the partial parsing can identify the phrases (for example noun-phrase or verb-phrase) or segments in the sentence. It can not analyse the phrases itself or assign what kind of relation the phrases have with each other within a sentence. However, it achieves higher accuracy than full parsing, and it can be carried out quickly and efficiently. Partial parsing might be applied as first step in full parsing as well.