Deep Learning-Driven Approaches

Software that works with human language is a field called natural language processing, often abbreviated NLP. For a long time, this field relied mostly on explicit rule-based systems and simpler statistical methods. Then Large Language Models, or LLMs, arrived and introduced a new family of approaches. Before reading on, consider the question below.

The source says that natural language processing previously relied mostly on two kinds of method. The first is explicit rule-based systems, in which the behavior of the software is defined by rules written out explicitly. The second is simpler statistical methods, which work from statistics gathered about language rather than from a set of hand-written rules alone. The source describes these two as the methods that mostly defined the field before the newer approaches arrived.

The important word in the source's description is mostly. The source does not claim that rules and statistics were the only methods in use, only that they were the main ones. This is a useful habit to carry forward: when a source describes a field's history, it is describing the dominant approaches, not listing every technique anyone ever tried.

A helpful way to compare approaches is to ask where each one gets its knowledge about language. Under explicit rule-based systems, that knowledge is written down by people as rules. Under simpler statistical methods, it is derived from counts and statistics gathered about language. Under the deep learning-driven approaches that LLMs introduced, the source describes a different starting point: the approaches advance understanding, generating, and translating language, and they are built around deep learning. The diagram below lays these three sources side by side and marks the shift between the earlier methods and the LLM-driven approaches.

Read the diagram from left to right. The two earlier methods on the left are the ones the source says mostly defined the field. The right side shows what the source attributes to LLMs: the deep learning-driven approaches they introduced, and the advances those approaches brought to understanding, generating, and translating language. The arrows show a shift in approach. They do not claim that every older method disappeared at once.

The source uses the word transformed to describe LLMs' role. LLMs transformed the field, which means the change is substantial. But the source is careful about what kind of change it is. The shift is a change in approach, not a change in the tasks the field addresses. A learner should therefore not picture LLMs as opening up an entirely new set of language problems. They are a new way of handling the problems natural language processing already cares about.

The source names three capabilities that deep learning-driven approaches advanced: understanding human language, generating human language, and translating human language. These three are the parts of the field the source identifies as changed by LLMs. The source does not describe a separate earlier method for each capability, so the story of the shift is the same for all three: earlier rules and statistics were the dominant approaches, and LLM-driven approaches advanced each capability.

The concept map shows the three capabilities as siblings. None of them is described as a step that leads to another, and the source does not rank them. What they share is the same source of advancement: the deep learning-driven approaches that LLMs introduced.