What are the core tasks of natural language processing?
A stable task list under changing techniques. Understanding tasks: classifying documents, extracting names, dates and amounts, judging tone through sentiment analysis, resolving what “it” refers to. Generation tasks: translation, summarization, answering questions, writing. Underneath both sit the mechanics, splitting text into units via tokenization and representing meaning as vectors. The task list is what defines NLP; the models that perform it have been replaced three times, from hand-written rules to statistical machine learning to neural networks.
How did large language models change NLP?
They collapsed the pipeline into a prompt. Classic NLP built one trained component per task and chained them; an LLM performs most of the task list from instructions alone, because pretraining on vast text taught it language structure the old pipelines encoded by hand. The architecture behind that leap is the transformer, an NLP invention that ate its own field. What remains of classic practice is economics and precision: specialized small models still win on high-volume, narrow tasks, where a prompted frontier model is too slow or too expensive per document.
What does NLP look like in a business today?
Like document work finally automated. The durable wins are intake and routing, reading invoices, contracts and applications into structured fields, triaging tickets and mail, monitoring reviews and support tone, and answering questions over internal documents. Language coverage matters more than demos suggest: performance is strongest in English, and a workload in Polish, German or Arabic deserves its own evaluation rather than an assumption. Evaluate NLP on your own documents and your own language, because published benchmarks measure neither your formats nor your edge cases.