If you don’t yet understand the word NLP, then this is the course to begin with. It does not assume that you’ve had some experience with machine learning in the past. You are presented with the jargon which will crop up in all of the other courses on this list, you get your first exposure to the common problems and uses of NLP, and an introduction to language models and deep learning as they relate to text. There’s a module dedicated to TextBlob, a very easy-to-use Python library, as well as sentiment analysis, and this little excursion into what is more of a computer vision topic actually becomes much more readable and understandable when it comes to language tasks once you see it. With a background in conversational AI and enterprise NLP deployments, you’ll sense how practically the material is presented by the instructor, Georg Huettenegger. More than 46,000 students have taken it and the rating is 4.5 stars – a pretty good endorsement for a free course. It will not make you as dangerous with transformers as the next person in a meeting who says, “tokenization,” but it will ensure that you do not feel lost the first time you hear the word in a meeting, and that’s enough.

Anyone who has been in the NLP community for any amount of time will have heard someone talk about this class. Pointed in the direction of not only using neural networks for language tasks, but of really understanding what is going on in them; it’s become something of a rite of passage, taught by Christopher Manning. It starts with word vectors, dependency parsing, and then progresses to self-attention, full transformer architectures, and the evaluation of LLM and the 5 programming assignments are really a challenge as they should be in a graduate course, and culminate in a final project in which students are able to apply a complete neural model to a large-scale NLP problem, sometimes successfully enough to publish. Whereas many of the “free university course” type offerings online are outdated, Stanford’s are, as they say, kept fresh, and recent courses are made available in the form of podcasts, lecture notes and problem sets, all of which are available to all for free. Plan your time for it. This is not a lunch break course!

The Hugging Face course is one of the best free courses to learn about transformers, and the tools they created are used by the majority of the industry. This relatively standard NLP course is now one of the single most useful and relevant LLM induction courses anywhere on the web, and it does so without any fuss or fanfare, as it covers the Transformers library, Datasets, Tokenizers, Accelerate and the Hugging Face Hub. The first couple of chapters introduce you to the concept of pipelines, how they work, and the difference between encoder, decoder, and sequence-to-sequence models, and then introduce fine-tuning using the Trainer API or Keras. The classic NLP tasks, datasets and tokenizers are covered in more detail in later chapters. You’ll move from know-how what a pipeline is to tuning your very own. Coding exercises are executed instantly in your browser, if you prefer to not install a neighborhood setting initially. It’s free, no ads and, importantly, Hugging Face says it’s ad-free.

This is not a single course, but rather a line of articles, and that is what it’s all about. Sitting through hours of structured video is not the way that everyone learns best, so this collection allows you to choose individual topics: linguistic resources; word-level and syntactic analysis; semantic analysis; word-sense disambiguation; discourse processing; POS tagging; information retrieval; general NLP applications; and Python-based implementations. The articles are independent of each other and there is no particular way they need to be read and no login wall prevents you from reading one article and closing the tab. This is a resource that you bookmark, not “complete”, but you use when you come upon a particular idea halfway through a project which you don’t quite understand yet.

While theory is great, this course is designed to fill the gap that comes in having a theory course. In lieu of another iteration of “here’s what tokenization means,” it turns NLP to problems that are more representative of how a business would use it: language detection, customer engagement analysis, social-media analytics, and a demo-heavy dive into the intersection of NLP and revenue growth and customer loyalty. A complete section with two demonstrations using NLTK, and a text-to-speech walk-through towards the end. It is taught by Sravan Malla, a senior AI/ML engineer and mentor for capstones at Great Learning, and is for someone who has the fundamental vocabulary already under their belt and is looking for how that translates to a business use case. Short at 3.75 hours, but more variety in the time than the duration implies.

One of those applications of NLP that everyone has experienced is machine translation, but very few people ever think about building it until they have to. The short course introduces the concept of sequence-based language processing through a step-by-step construction of a working example translating English to French, instead of being dropped into a ready-made black box. It’s small, and rather than a thorough technical research overview of neural machine translation, it’s a solid overview of how sequence modeling relates to a task you already sense intuitively. This works well for people who have learned the fundamentals of NLP before and do not want to skim the surface of an application but rather see it done correctly.

Whether it’s marking spam, directing customer service tickets or determining if a review is good or bad, text classification is one of the most commonly used NLP applications in the real world. Independent consultant and DataCamp instructor Sowmya Vivek will take students through what text classification really is and how it works and then dive into a full walkthrough of the three key modeling approaches: Naive Bayes, linear classifiers, and support vector classifiers, using Python. This is a course for beginners, but it’s not a purely conceptual course. Models will not just be described but seen in action and tested. It has a high rating of 4.6 from a few thousand learners, and its reviewers particularly comment on how theory is integrated with hands-on work in Python, rather than being two halves of a course that are bolted together.

This is one that’s at the nexus point of classic NLP and the generative AI tsunami that has consumed most of the discussion about language tech lately. It’s short, it focuses primarily on the task concept itself, the pre-processing of this task, the feature extraction for this task and how to use ChatGPT to actually write code that is capable of performing a task like summarization, instead of just talking about the idea in abstract. It does NOT take the place of a prompt engineering course, but it is a quick and effective way to get from the basics of NLP that you already know to the generative tools that everyone is using today—and it does it in less than an hour!

Mostly due to the fact that it abstracts a lot of the complexity that libraries like spaCy or NLTK reveal to the user, TextBlob is one of the most user-friendly ways to get into python based NLP. This course begins by covering the basics of NLP related to TextBlob specifically, stopwords, lemmatisation, why and how to clean text data and the different methods of text vectorisation, before progressing on to the functionality provided by the library and a complete use case for sentiment analysis using TextBlob and Python. If you like to have some hands-on coding experience before diving into more complex and configuration-heavy tools such as spaCy’s pipeline system, it’s a good stepping stone.

Large language models are the components of NLP that people outside the profession know, so it’s logical to have a short, structured introduction here. The course covers five modules: an overview of what an LLM really is, the languages and systems they operate on, including some popular examples, how transformer architecture influences their processing and generation capabilities, some real business case examples of when they’ve been used, such as creating product descriptions or debugging code, and last but not least, challenges and limitations, hallucinations, bias, domain mismatch, and scalability, and where the field is going next. It is very much targeted towards the beginner and reviewers repeatedly note that the progression from earlier NLP to today’s LLMs gets explained very clearly, and without requiring you to have read a research paper on attention.

Large language models are the components of NLP that people outside the profession know, so it’s logical to have a short, structured introduction here. The course covers five modules: an overview of what an LLM really is, the languages and systems they operate on, including some popular examples, how transformer architecture influences their processing and generation capabilities, some real business case examples of when they’ve been used, such as creating product descriptions or debugging code, and last but not least, challenges and limitations, hallucinations, bias, domain mismatch, and scalability, and where the field is going next. It is very much targeted towards the beginner and reviewers repeatedly note that the progression from earlier NLP to today’s LLMs gets explained very clearly, and without requiring you to have read a research paper on attention.

This book is much more comprehensive than the library documentation, written by the library’s very own developers, Steven Bird, Ewan Klein, and Edward Loper. It covers such topics as accessing real text corpora, processing raw text, writing structured programs, categorizing and tagging words, learning to classify text, extracting information, analyzing sentence structure, building feature-based grammars, and reasoning about the meaning of sentences—and it does so by using working code and real data, not toy examples developed for a slide deck. Its final chapter is a more philosophizing look at the problem of language in general, rather than using code—this is a nice, thoughtfully done book that ends that way! Some places it is dense, some places it is more like a proper textbook than a quick tutorial, but it’s the kind of resource that you leave open in a browser tab for months, as you work on real projects on top of it.

When building spaCy, we had production use in mind and our official course, developed by Ines Montani, one of the core developers of spaCy, and co-founder of Explosion, shows that in every part of it. Developed on an interactive, browser-based format that allows you to run real code, and to build full pipelines, train your own models, and use custom components, all without having to install anything first. It is broken up into four chapters that progress from the very basic to the very advanced in pipeline modding. Mostly because you’re not sitting in a pile of slides for the first one or two minutes before you pick up a pencil and start writing real working code, it’s one of the more genuinely fun free courses on this list.

This will run in about 3 hours as a single, long-form video, developed by Dr. W.J.B. Mattingly, a postdoctoral fellow at the Smithsonian Institution’s Data Science Lab, and is a nice complement to the official spaCy course if you prefer to learn watching someone build something together rather than reading a doc alone. This includes installing spaCy, using spaCy’s data containers, linguistic annotations, named entity recognition, word vectors, building pipelines, the EntityRuler and Matcher components, building their own components, and using a lot of regular expression to deal with simple patterns and multi-word tokens; and concludes with an example of building named entity recognition for a financial domain. There are also complementary jupyter notebooks provided with the video, which is a very nice addition if you want to play along and do it yourself instead of sitting with someone else’s shoulders to the keyboard for three hours.

This is the specialization that integrates the following four items on this list, and if you are the kind of person who would prefer a multi-month course over one that is a random assortment of one-off classes, this may be the single best free option on this list. It was built and taught by Younes Bensouda Mourri, an AI instructor at Stanford who helped develop the original Deep Learning Specialization, as well as Staff Research Scientist at Google Brain, Łukasz Kaiser, who co-wrote the original Transformer paper, and brings you from sentiment analysis and vector spaces all the way up to sequence models and attention mechanisms, culminating somewhere between where the field of modern transformer-based NLP sits today. Rated 4.6 stars with over six thousand reviews, it runs about three months at 10 hours a week (with suggested pace) and each course within it can be audited for free, and a certificate of completion is available for a fee if desired.

A good single course to take if you are primarily interested in the text classification side of things and not the four-course sequence above. You’ll learn logistic regression and Naive Bayes, building a tweet sentiment classifier from scratch, along with extracting numerical features from the raw text; you’ll learn about word frequencies and get some insight into the semantic relationships between words; and then you’ll learn about dimensionality reduction with PCA and actually visualize the semantic relationships between words. The last module is on locality-sensitive hashing for machine translation and document searching—a smart technique that is not taught in most introductory courses. It’s a course that makes NLP feel like a black box, as you’re creating the classifiers yourself, not importing a pre-trained one and hoping it works out.

Course two of the specialization, and perhaps the part most people would skip to move on to the transformers, is a little of a mistake. Despite the modernism of attention models, the course takes a retro approach to sequence prediction by teaching useful intuitions from HMMs and n-gram models, and while we’re at it, there’ll be some practical examples for your use, such as autocorrect and autocomplete, as well as a dedicated chapter on word embededdings when viewed as a probabilistic model and not just a neural one. It helps to understand the rationale for their success, particularly where they failed, so that one can better appreciate why neural sequence models eventually took over the field that they have.

In course three, the course will dive into recurrent neural networks, gated recurrent units, LSTMs and Siamese networks and use them in all the modules of the course to process sentiment analysis, text generation and named entity recognition. This is where the material really feels modern, as these are the predominant architectures for sequential language data up until attention mechanisms and transformers pretty much took over the entire field. The probabilistic foundations from course 2 and the attention-based material from course 4 (which typically uses the Siamese network module in its application to sentence similarity) is a useful bridge between those two courses and is not covered nearly as often in competing free courses.

The last course of the specialization, and the one that finally takes you all the way to transformers. In its first modules, it introduces attention mechanisms, encoder-decoder architectures, and neural machine translation, followed by a comparison of RNNs and sequential models with the more recent Transformer architecture, and a walkthrough of creating a text summarization model, before ending with transfer learning using cutting-edge models such as T5 and BERT to create a working question-answering system. It is the same concept you will need to understand BERT, GPT models, and basically anything new that has been created since. This is where the pay-off really begins if you have completed the first three courses successfully.

Upfront, I should state that this is the same course that’s ranked No.3 – Hugging Face doesn’t offer a standalone “Transformers Course” as such. If you’re familiar with the NLP Course above, you’ve already covered this material, in the chapters covering transformer architecture, working with pretrained models, fine-tuning workflows, model evaluation, and Hugging Face Hub. It remains here because there are many people who come to Hugging Face’s ecosystem and the reason is transformer training and not just general NLP and it is really helpful for them to know that they end up in the same place.

Despite the dominance of newer architectures in the headlines now, BERT, which stands for Bidirectional Encoder Representations from Transformers, was the model that showed that transformers can learn rich, bidirectional language representations that can process text in either direction and is still worth learning. This is not a typical course with a syllabus, but it works as if it were due to the checkpoints of the pretrained BERT models that can be loaded directly, the details of its WordPiece tokenization scheme, model input format, full fine-tuning workflows, and how to use BERT for tasks like text classification and token classification with working code examples throughout. It isn’t as “hand-holdy” as a dedicated course, so search a bit on your own, but as a hands-on guide to one of the most fundamental of architectural structures in the field, this one has plenty of use.

In many NLP courses, tokenization is just a side note, which is a mistake as it directly influences the understanding of the transformer model for everything it consumes. This tutorial shows how to build a tokenizer from scratch with Byte-Pair Encoding, train it in seconds on a realistic data set, such as WikiText-103, and explains subword tokenization in greater detail, attention masks, and how a trained tokenizer seamlessly integrates with a pretrained model further downstream. It’s a smaller and more specific text than may be on this list, but comprehending it correctly later makes the bizarre behavior of a model for so many reasons less mysterious. A lot of “the model is behaving weirdly” bugs stem from the initial tokenization of a text, and it is quite surprising how many there are that fall into this category.

The surprising number of practical NLP applications include customer analytics dashboards, legal document processing, news aggregation, and more: they’re all backed by named entity recognition, which is the ability to extract people, organizations, dates, locations and other predefined categories from unstructured text. It covers how to use spaCy’s out-of-the-box, pretrained statistical models for entity extraction, how entity labels are structured, how they are visualized, and most importantly, how to tailor or retrain those models to work for your own domain, which is especially crucial when dealing with data outside the scope of general news or generic web text, such as in medical records or financial filings, where common models are unlikely to catch entities they weren’t trained for.

For all these reasons, sentiment analysis is often the initial “real” NLP application that most folks build themselves, and it makes sense: it’s easy to explain to someone else, easy to sanity-check with your own understanding of a sentence, and a walk-through of the entire pipeline from raw text to working prediction in one easy project. The material builds incrementally from simple rule-based scoring methods to more powerful supervised machine learning techniques, using real language (such as product reviews, social posts, etc.) and not sanitized textbook sentences that never look much like your real-life text inputs later in the book.

To reiterate: This is not a standalone resource, but is part of courses 3 and 20, which are project-oriented and include fine-tuning-focused late chapters of the same Hugging Face course. However, it is definitely worth considering as a standalone task, such as building something end-to-end, sentiment analysis, text classification, question answering, or summarization, using a pretrained transformer. Typical workflows are: (1) load a raw text dataset, (2) preprocess the data, (3) tokenize the data, (4) load a transformer model, (5) fine-tune on your task, (6) evaluate results, and (7) wrap in a usable application. Once you’re familiar with the basic loop, you can experiment with the following projects: Customer review sentiment analysis, Spam Message Detector, News Article Classifier, Resume Information Extractor, or a simple Question Answering system with a small, well-defined knowledge-base under your control.