SSkilvy
RULog inStart free

A spectrum of tasks: from simple analysis to generation

NLP is not one task but a whole spectrum. Understanding the task types helps choose a tool and correctly formulate the problem. Let us cover the main groups: some analyze text (what is in it), others extract data, others create new text. Many real systems combine several tasks.

Classification sentiment, spam, topic Extraction names, dates, amounts Generation summaries, answers, text Search by meaning, similar
NLP tasks: classification, extraction, search, translation, summarization, answering, generation.

Groups of NLP tasks

  • Classification: determine a texts category. Review sentiment, spam/not spam, ticket topic, intent.
  • Information extraction: pull data from text. Names, dates, amounts, organizations (entities), key facts.
  • Search and similarity: find the relevant by meaning (embeddings from the vector-search course). Smart search, the similar, duplicates.
  • Generation and transformation: create/transform text. Translation, summarization, question answering, generation.
Cover the main NLP tasks with application examples.
The main NLP tasks by group, with applications: 1) ANALYSIS/CLASSIFICATION (understand what the text is): — Sentiment: positive/negative/neutral. Application: analyzing reviews, social media, feedback — what customers feel. — Topic/category classification: spam/not spam, support-ticket category, document type, article topic. Application: email sorting, ticket routing. — Intent detection: what the user wants (order, check status, complain). Application: chatbots, assistants. — Language, toxicity detection etc. 2) INFORMATION EXTRACTION (pull data): — Named-entity recognition (NER): find names, organizations, dates, places, amounts, numbers in text. Application: document processing, extracting data from emails/contracts, autofill. — Relation/fact extraction: not only "there is a date" but "this date is the payment deadline". — Keyword/phrase extraction. 3) SEARCH AND SIMILARITY (find the relevant): — Semantic search: find texts by MEANING (embeddings, from the vector-search course). Application: smart knowledge-base search, RAG. — Finding the similar, duplicates, text clustering. 4) GENERATION AND TRANSFORMATION (create/change text): — Machine translation: language to language. — Summarization: compress a long text into a brief one. — Question answering: over text/a base (a link to RAG). — Text generation: create coherent text (powerful in LLMs) — write an email, answer, description. — Paraphrasing, correcting, simplifying text. 5) SPEECH (at the audio intersection): — Speech recognition (speech to text) and synthesis (text to speech) — a link to the voice/multimodality course. Key observations: — Tasks of different complexity: classification is simpler, generation harder (creating good text is harder than classifying). — Real systems COMBINE tasks: e.g. a support chatbot = intent detection (classification) + data extraction (entities) + base search (semantic) + answer generation. Or review analysis = sentiment classification + extraction of mentioned topics + summarization. — LLMs unite much: modern large language models can solve MANY of these tasks with one model via a prompt (classify, extract, summarize, translate, answer, generate). This changed NLP: previously each task had its model, now often one LLM flexibly solves different things (a link to the LLM courses). The practical meaning of the choice: — Determine which TASK (or combination) solves your problem: are you analyzing text? extracting data? searching? generating? — Choose the tool for the task: for simple classification/extraction at volume a specialized model may suffice (more accurate/cheaper); for flexible tasks, generation, combinations — an LLM (as in the specialized vs multimodal choice in CV). — A combination of tasks and tools is often optimal. Practical rule: NLP is a spectrum of tasks (classification, extraction, search, generation, speech). Determine the needed task or their combination for your problem. Modern LLMs solve many tasks with one model via a prompt, but for narrow repeated tasks at volume specialized models are sometimes more accurate and cheaper. Understanding the task map helps correctly formulate the problem and choose the tool.

Real systems combine tasks

An important observation: real NLP systems usually combine several tasks. Example: a support chatbot = intent detection (classification) + data extraction (entities) + base search (semantic) + answer generation. Review analysis = sentiment classification + extraction of mentioned topics + summarization. Understanding individual tasks as "building blocks" helps assemble solutions for a specific need. Determine which task or their combination solves your problem: are you analyzing text? extracting data? searching? generating?

LLMs unite much

A key shift: modern LLMs solve many tasks with one model via a prompt — classify, extract, summarize, translate, answer, generate. Previously each task had its model, now often one LLM flexibly solves different things (a link to the LLM courses). But, as in CV (multimodal vs specialized), for narrow repeated tasks at volume specialized models are sometimes more accurate and cheaper, while for flexible tasks and generation an LLM is more convenient. Understanding the task map helps correctly formulate the problem and choose the tool.

Rule: NLP is a spectrum of tasks (classification, extraction, search, generation, speech). Determine the needed task or combination for the problem. LLMs solve much with one model via a prompt; for narrow tasks at volume specialized models are sometimes more accurate and cheaper.

🧠 How did modern LLMs change solving NLP tasks?

← BackNext →