[Podcast] How machines deal with human language?

Written by Alana Team
on July 27, 2020

Now that you know if machines think like humans, the time has come to understand how they deal with human language and what is the current ability to understand how we communicate.

In this episode, Marcellus Amadeus, CTO of Alana AI, talks about the language limitations of machines and explains how they use mathematical tricks to understand humans.

Human Language and Communication

To understand how machines deal with human language, we first need to talk about the concept of language and its role in communication. According to the dictionary, the definition of the concepts is:

  • Language is any systematic means of communicating ideas or feelings through conventional signs, sounds, graphics, gestures, etc.
  • Communication is the action of transmitting a message and, eventually, receiving another message in response

For communication to exist, the interpretation of language must happen. This point is key to both human and machine communication.

For example, imagine traveling to a country that you do not speak the language. You will not be able to interpret the signs of the local language, you may understand some universal image, but communication will be flawed.

The same is true with machines. If they are unable to interpret human natural language, they cannot communicate with quality. For that to happen, there is the NLP (Natural Language Processing) method

Natural Language Processing (NLP)

Natural Language Processing (NLP) is a sub-area of ​​computer science, artificial intelligence, and linguistics.

Within artificial intelligence, it is an essential area in terms of ensuring the quality of communication with people, since NLP is the technology used to help computers understand human language.

It studies the problems of generation and understanding of human languages ​​and has been evolving since the 1950s. We can divide this evolution into three phases:


Evolution of NLP

The symbolic approach is based on lexical rules, those developed by man, that is, it follows mandatory speech, which is materialized and registered by specialists so that the systems can follow them.

The second phase, which is statistical, is based on observable and recurrent samples of linguistic phenomena. For example:

  • The models recognize recurring themes through mathematical analysis of large parts of the text; 
  • When identifying a trend, the computer system can develop its own language rules that will be used to analyze future inputs and / or the generation of language output.

Finally, the neural approach, which follows supervised learning and is entirely numerical. This is the model we use today. It works as follows:

  • First, it turns the sentence, which is language, into numbers (vectors;
  • The sentence is now represented by the numbers assigned to each word, so it can be used by neural network algorithms;
  • The neural network algorithm analyzes and finds patterns in the numbers for interpretation.

Curiosity about NLP

Marcellus took the opportunity to share a very interesting historical fact.

NLP systems were one of the first applications of artificial intelligence. In the 1950s an application for translating from Russian to English was created and it was based on a bilingual dictionary and had predefined rules.

Another famous historical application based on NLP is the

ELIZA chatbot, created in 1965, which was one of the first chatbots in history to be able to perform the Turing test.

Imagem da tela de um computador demonstrando o chatbot ELIZA

Limit in communication between machines and humans

One of the main challenges in developing an AI is dealing with the question: how does the machine interpret the context? 

Marcellus says that machines are getting smarter, but there is a crucial limiting factor that is the context! Despite being able to interpret symbols, machines cannot interpret the context itself.

Some more modern algorithms, for example, generate text and interpret it accurately and consistently, but the machine does not genuinely feel something.

 Scalable technology

People involved in the development of artificial intelligence are looking for ways to make technology more scalable by increasing the machine's ability to understand.

For example, allowing the machine to learn from information it already has available, creating data to feed itself.

In general, advances in NLP are fundamental to improve the interactions of artificial intelligence with humans.

Want to know more about artificial intelligence? Follow the Inside Alana Podcast and listen to available episodes, whenever you want.

personagens escrevendo




Você também pode gostar:


[Podcast] The challenges of artificial intelligence in communication

For many, artificial intelligence exists only in films or conspiracy theories about the end of the world. Although it is...


Alana AI expands operations in Latin America

Founded in 2015, Alana AI is opening offices in Mexico and Puerto Rico. Today, the artificial intelligence company alrea...


What changes with the General Data Protection Law (GDPL) in Brazil?

Did you know that the General Data Protection Law was enacted in September this year and will take effect from the last ...

Receba insights sobre
Inteligência Artificial por e-mail,
na medida certa.