The cournerstone of any natural language understanding system is NLP algorithms. The classical set of algorithms constitutes an NLP ptocessing pipeline:

Limitations of syntactic approaches:

Consider the task of classification, where you solve a problem of mapping a set of features \(X\) to a target label \(y\), so that \(C(X)=y\), where \(C\) is your classification function. Now, it is highly possible that your set of features does not provide a perfect explanation of the target class, and thus you may find several data samples that are identical, but have different class labels, i.e. \(X_1=X_2, C(X_1)=y_1, C(X_2)=y_2\). This is usually called “noisy data”. The process of training a model condescends to finding a function \(C\) that classifies correctly, but what is the measure of correctness, especially in the presence of the noise?

Motivation