Structured Prediction

One of the simplest and easiest ways to talk about structured prediction is that it uses training problems to solve a classification task. A resource available from NeurIPS quoted by Sasha Rush in July of 2010 describes it as: “a framework for solving problems of classification or regression in which the output variables are mutually dependent or constrained.”

Specifically, when a prediction cannot be solved by direct observation of all possible values, the structured prediction takes inputs, and uses them to predict the results.

Alexander Passos, then a PhD ML student at UNICAMP in Brazil, gives an interesting definition of structure prediction in Quora that is abundantly useful in characterizing this sort of utility: “Structured prediction is a special case of multi class classification (that is, given x predict y) where:

1. There are far too many possible values for y (exponential or infinite).
2. However, these values are not opaque, and inspecting their structure can help you design a classifier that learns from few examples (in relation to the cardinality of y) in a short amount of time.”

Structured prediction has been useful in natural language processing, bioscience research and other disciplines. For instance, using sequence tagging and parse trees, a structure prediction program can achieve various natural language processing goals.