After this lesson, you will be able to: Map the modern NLP research landscape, the tasks the field works on, and what 'research-ready' means, so you know what you are building toward.
Before the architecture, get oriented. This lesson surveys what NLP research actually studies in the LLM era, the venues and how they work, and the shape of a research contribution, so the technical lessons that follow have a purpose.
This is a free introductory lesson. No purchase required.
Classical NLP split into many tasks: classification, named-entity recognition, parsing, machine translation, summarization, question answering. LLMs collapsed many of these into one interface (prompt in, text out), so research shifted. Active areas today include: efficiency (smaller/faster models, quantization, distillation), evaluation (the field admits it cannot measure its own systems well), alignment and safety, reasoning, multilinguality and low-resource languages, interpretability (what is the model actually doing inside), retrieval and long context, and data (what to train on, contamination, synthetic data). A workshop paper usually makes a small, sharp contribution to one of these.
The *ACL (the Association for Computational Linguistics)* runs the field's main conferences: ACL, EMNLP, NAACL, EACL, plus the journal TACL and the curated CL journal. Each conference hosts a Student Research Workshop. Submissions now flow through ACL Rolling Review (ARR): you submit to ARR, get reviews on a monthly cycle, then 'commit' a reviewed paper to a specific venue. Deadlines anchor the year, so researchers plan projects around the ARR cycle and the conference dates. The ACL Anthology hosts every accepted paper, free, forever, which is your primary reading source.
A paper is not 'I built a thing.' It is 'I asked a precise question, ran an experiment that answers it, and the answer is interesting and trustworthy.' Common contribution types at SRW scale: a new method that beats a baseline on a task, a careful analysis that reveals something surprising about existing models, a new dataset or benchmark, or a reproduction that confirms or challenges a prior claim. The unit of credibility is the experiment plus the evaluation, not the size of the system.