Lesson 4 — The tasks, and which ones work
NLP is not one problem. It is a family of tasks with very different maturity levels, and confusing a solved one with an unsolved one is how projects get committed to on the strength of a demonstration.
Classification
The task: assign a label to a piece of text. Which department should handle this ticket? Is this review positive? Is this email spam? Is this document a contract, an invoice or a letter?
Maturity: solved. This is the most reliable task in NLP and the workhorse of production systems. A fine-tuned pre-trained model reaches high accuracy with a few hundred to a few thousand labelled examples per class.
The trap: your label definitions. Teams routinely discover that two annotators disagree on 20% of examples, which caps the model's achievable accuracy at the level of human agreement. If people cannot agree what "urgent" means, no model will predict it.
Sentiment analysis, specifically
Worth separating out because it is oversold. Classifying clearly positive against clearly negative reviews works well. What works much less well:
- Sarcasm. "Brilliant, another delay." Positive vocabulary, negative meaning.
- Mixed sentiment. "Great camera, terrible battery." One label cannot represent this, which is why aspect-based sentiment analysis exists.
- Domain shift. A model trained on film reviews performs poorly on medical feedback, where the vocabulary of severity is entirely different.
- Neutrality. Most real text is neither positive nor negative, and forcing a binary choice manufactures signal that is not there.
Named entity recognition
The task: find and label the entities in text. People, organisations, places, dates, amounts, product codes, drug names.
Maturity: solved for common types, needs work for your own. Pre-trained models handle people, places and organisations well. Extracting your specific entities — internal reference formats, industry-specific product names, clause types in contracts — requires annotated examples, and this is where document-processing projects spend their time.
It is also the foundation of most document automation: pulling fields out of invoices, contracts and forms is named entity recognition wearing a business label.
Machine translation
Maturity: excellent between well-resourced languages, poor otherwise. Translation between English, French, Spanish, German, Chinese and a dozen others is good enough for most purposes, and this is the field's most visible success — transformers were invented for it.
The honest caveats:
- Low-resource languages are far behind, sometimes drastically. The gap tracks the amount of digitised text available, which tracks economic and colonial history.
- Domain terminology needs adaptation. A general model mistranslates legal and medical terms with complete confidence.
- Cultural and idiomatic content flattens. Humour, register and politeness conventions do not survive intact.
- Consistency is not guaranteed. The same term may be rendered differently across a long document, which matters in technical documentation and legal text.
Summarisation
Two genuinely different tasks that get one name.
Extractive summarisation selects the most important existing sentences. It cannot invent anything, so it is safe, and it reads choppily.
Abstractive summarisation writes new sentences. It reads far better and it can fabricate, including facts and figures absent from the source.
Maturity: use with review. The output is fluent and confident, and fluency is not accuracy. For any summary that will be acted on — medical notes, legal documents, financial reports — a human must check it against the source. Extractive is the safer default when the stakes are high.
Question answering
Two variants with very different reliability.
Extractive: find the answer inside a provided document. Reliable, because the answer is quoted rather than composed, and verifiable by construction.
Generative: produce an answer from the model's own training. Unreliable for facts, because the model has no way to distinguish what it knows from what it is inventing.
The engineering answer is retrieval-augmented generation: retrieve the relevant passages first, then let the model answer using them, with citations. This is why RAG became the standard pattern for putting language models to work on a company's own documents, and the premium RAG course covers building one.
Speech
Speech to text: solved for clear audio in major languages. Transcription crossed the usability threshold around 2017 and has kept improving. It degrades on overlapping speakers, heavy background noise, strong accents under-represented in training data, and specialist terminology.
Text to speech: solved to the point of raising other questions. Synthetic voices are now hard to distinguish from recordings, and voice cloning from a short sample is widely available — which turns a technical achievement into an impersonation risk.
The honest scoreboard
| Task | Status | Needs human review? |
|---|---|---|
| Text classification | reliable | spot checks |
| Named entity recognition, common types | reliable | spot checks |
| Named entity recognition, custom types | needs your annotated data | yes, while building |
| Translation, major language pairs | reliable | for published material |
| Translation, low-resource languages | weak | always |
| Extractive summarisation | reliable | light |
| Abstractive summarisation | fluent, may fabricate | always |
| Extractive question answering | reliable | light |
| Generative question answering without retrieval | unreliable for facts | always |
| Speech transcription, clear audio | reliable | light |
| Sentiment, clear cases | reliable | light |
| Sentiment, sarcasm and mixed | weak | yes |
Extractive tasks quote; generative tasks compose. Anything extractive can be checked against its source, which makes it safe to automate. Anything generative can invent, which makes review mandatory wherever the output is acted upon. Choosing extractive where you can is the single most useful design decision in an NLP product.
In three sentences
Classification, named entity recognition for common types, translation between well-resourced languages, extractive summarisation and speech transcription are all reliable enough to build on. Abstractive summarisation and generative question answering are fluent and can fabricate, so they require human review wherever their output is acted upon. The most useful design distinction is extractive versus generative: extractive output can be verified against a source, which is what makes automating it defensible.