
Natural language processing (NLP) is with no doubt – in my opinion – the most famous field of Data Science. Over the past decade, it has gained a lot of traction "buzz" in both industry and academia.
But, the truth is, NLP is not a new field at all. The human desire for computers to comprehend and understand our language has been there since the creation of computers. Yes, those old computers that could barely run multiple programs at the same time, nevertheless comprehend the complexity of natural languages!
Natural language processing – if you’re new to the field – is basically any human language, such as English, Arabic, Spanish, etc. The difficulty behind giving computers the ability to understand natural languages is how complex they can get.
When we talk, we often enunciate words differently; we could have different accents regardless of whether we are using our mother language or another. We also often tend to mix words together when we talk to reach our points faster. Not to mention all the slang words that come up every day.
The purpose of this article is to shed some light on the history of NLP and what are the subfields of it.
How did NLP start?
NLP is an interdisciplinary field that combines both computer science and linguistics. Let’s – for the sake of this article – consider the English language. There is an infinite number of ways we put words together to form a sentence. Of course, not all these sentences will be grammatically correct or even make sense.
As humans, we have the ability to distinguish between them, but a computer can’t. Moreover, it’s illogical to give the computer a dictionary with every possible sentence in all possible languages!
So what should we do?
In early NLP, scientists proposed dividing any sentence into a collection of words that can be individually processed is much easier than processing the sentence as a whole. This approach is similar to how we have been taught language as children or adults learning a new language.
When we are first introduced to a language, we are first taught the parts of this language’s speech. For example, there are 9 main parts of speech in English, like nouns, verbs, adjectives, adverbs, pronouns, articles, etc. These parts of speech help us understand the purpose of each word in the sentence.
Knowing the word category is not enough, especially for words that could have more than one meaning. For example, the word "leaves" could be a verb meaning to leave or the leaf’s plural form.
Because of that, computers needed to know a little bit of grammar and help it if they got confused about a certain word. And here where phase structure rules were established.
In short, these rules are a collection of grammar laws that forms a sentence. In English, a sentence can be formed by combing a verb clause with a noun clause. For example, She ate the apple. Here, "she" is the noun clause, while "ate the apple" is the verb clause.
The different clauses can be formed using different structures. With more phrase structure rules, we can create a parse tree to categorize every word in a specific sentence and eventually reach the overall sentence’s meaning.

This is all great, as long as our sentences are simple and clear, but the problem is, sentences can get really complex, or some may use not very con words – Shakespearean-like – to communicate their ideas. In this case, computers will find it difficult to understand what we mean.
The different subfields of NLP
Text Processing
Chatbots are one of the well-known examples of NLP. In the early stages of NLP, chatbots were rule-based. That meant scientists needed to code hundreds – maybe thousands – of phrase rules to map what the human types in with what the chatbot should reply. An example of that appeared in the ’60s. It was a therapist chatbot called Eliza.
Today, most chatbots and virtual assistants are built and programmed using different Machine Learning techniques. These machine learning techniques depend on multiple gigabytes of data collected from human-to-human conversations.
The more the amount of data supplied to the machine learning model, the better the chatbot will get.
Speech Recognition
So, chatbots are how computers understand written language, but what if the language was spoken? How can computers turn sound into words and then understand their meaning?
This is the second subfield of NLP, speech recognition. Again, speech recognition is not a new technology at all. In fact, it has been the focus of many researchers for the past decade. More precisely, in 1971, Harpy was developed at Carnegie Mellon University. Harpy was the first computer program to understand exactly 1000 words.
Back then, computers weren’t powerful enough to handle real-time speech recognition – unless you’re talking really slowly. This obstacle has been overcome with the development of faster and better computers.
Speech Synthesis
Speech Synthesis is very much the opposite of speech recognition. It’s giving the computer the ability to make sounds – or say words.
In speech recognition and chatbots, sentences are broken down into words, or as they are called in linguistics phonetics. These photonics can be stored and then rearranged and played back by the computer to say a specific sentence.
The first-ever speech synthesis machine was proposed in1937 by Bell Labs, and it was hand-operated. This changed over the years. The way phonetics was collected and put together was – and still is – the reason when computers "talk", it sounds robotic.
This roboticness has gotten better, but the use of modern algorithms, the latest virtual assistants, like, Siri, Cortana, and Alexa are proof of how far we have come. Yet, they still don’t sound fully human.
Takeaways
Natural language processing is one of the most famous data Science fields, and it is also one of the most important ones. In short, NLP is giving computers the ability to understand and produce human languages.
NLP is an umbrella name that covers many subfields and applications, but we can categorize them into three main categories, text processing, speech recognition, and speech synthesis.
All these categories and NLP applications often build upon machine learning models -mainly, neural networks – and a ton of human-to-human conversations. But, most often, techniques and grammar sets are for specific languages. Hence, their accuracy may not be very high, especially if we are dealing with different accents.
That’s why specific language models are often applied to any of these categories to increase their accuracy and make them more "human."
This article is the first in a series I plan to write that covers different aspects of NLP. Starting from its history, a brief introduction of each category, how to get started, some applications, and the current research status.
So, stay tuned!