Natural Language Processing
Definition
natural language processing (NLP) refers to the technology, or the discipline studying it, that enables computers to understand, interpret, and process human language. It is commonly called by its English abbreviation, NLP.
Explanation1
natural language refers to a language that people use in everyday life, such as Korean or English, and it is a concept contrasted with a formal language, which is artificially designed according to strict rules like a programming language. Representative tasks of natural language processing include the following.
- machine translation: transferring a sentence in one language into a sentence in another language
- text classification: dividing documents into predetermined categories. Spam mail filtering, sentiment analysis, and the like belong here.
- question answering: taking a question as input and producing an answer to it
- language modeling: estimating the probability distribution of the word that follows a given context. It forms the foundation of large language models (LLM).
The reason natural language is tricky to handle is that it is unstructured data of varying length with no fixed frame, and it is a sequence of discrete symbols—characters and words—so that operations such as addition or multiplication are not defined on it as such. A neural network is ultimately a function that takes a vector as input and produces a vector, so in order to handle natural language with a neural network, one must first convert text into numerical data. Thus, today’s natural language processing generally follows the following pipeline.
$$ \text{텍스트} \overset{\text{토큰화}}{\longrightarrow} \text{토큰열} \overset{\text{임베딩}}{\longrightarrow} \mathbb{R}^{d}\text{의 벡터열} \overset{\text{모델}}{\longrightarrow} \text{출력} $$
First, the text is split into small units called tokens, and each token is converted into a vector through an embedding. This sequence of vectors so obtained is fed into a model such as a recurrent neural network or a transformer to obtain the desired output, such as a translation, a classification result, or the probability distribution of the next word.
Historically, early natural language processing was dominated by rule-based methods that relied on grammar rules and dictionaries hand-written by linguists. In the 1990s, statistical methods that estimate the occurrence probability of words from a corpus replaced these, and in the 2010s, neural-network-based machine translation using recurrent neural networks, LSTMs, and the like produced good results and became the standard. Afterward the transformer (2017), an architecture composed of attention, appeared, and large language models trained on huge corpora are at the center of natural language processing at present (as of 2026).