Feature structure is used to bundle together phonemes and to give them binary values. The idea was first proposed by Roman Jacobson in 1949. It is used as a part of grammar studies, particularly studies into the structure of language and grammatical interactions. In essence, a feature structure is an attempt to understand the structure of grammar and make a workable system for representing how different elements of grammar interact.
Each feature structure is represented using an attribute-value matrix (AVM). This is a set of terms and symbols contained within brackets. It almost looks like the grammatical equivalent of a mathematical formula.
AVMs are divided into two columns. The left-hand column represents the features such as category and agreement. The right-hand column represents the sub-features of each feature such as gender and number. Each sub-feature in the right-hand column is attributed a value. For gender, the value would be male, female or neuter, and for number it would be single or plural.
An alternative method for representing a feature structure is a directed acyclic graph (DAG). The graph begins with a dot, also called a node. Each feature diverges off from the node along a curved arrow. Each feature can then split into sub-features. The sub-features end in another node that contains the sub-feature’s value.
An example of this would be a node leading to the agreement feature that leads to the person sub-feature and a value of 3rd. This represents a sentence told in the third person. In languages where gender is an important feature, the feature arrow could split into gender as well as person and would result in two values at the end. For example, this could mean third person and female.
Unification in a feature structure means that two features can split into sub-features, which then merge. Mergers within unified feature strands have the same value and are best represented using a DAG rather than an AVM. As well as unifying feature structures, attempts at unification can also prove certain structures are incompatible.
Feature structures are used for text encoding initiatives (TEI). These create mark-up schemes for linguists. Such schemes can then be used to analyze or interpret encrypted and encoded texts.
There are two main problems with using a feature structure. Firstly, feature structures produce a lot of generalizations. Secondly, they and unification structures cannot contain all possible values in a language.
A suggested solution to this problem is to include a third column or master branch called types. Each type would organize features into appropriate sections or classes. By doing this, the features would be regulated and so would the values each feature can take. The types would work in a hierarchy system to regulate feature interactions.