Having worked in the field of computer science as a professor and researcher since the 1980s, Maria das Graças Volpe Nunes knows what the latest artificial intelligence (AI) programs can and cannot do, seeing through the hype currently surrounding the topic. In this interview, given via video call, she exposes the most common misconceptions related to AI, such as the belief that machines are capable of understanding our questions, when actually they only provide the most likely sequences of words in response to those that were presented.
She also warns about the lack of questions being asked about the scope and possible errors of these programs. “If something goes wrong, who should be held responsible: the machine, the programmer, or whoever fed the machine with its underlying data?” asks Nunes, 65, who spent her career at the Institute of Mathematical and Computer Sciences of the University of São Paulo (ICMC-USP) in São Carlos, where she continues to work as a senior professor.
The researcher dedicated herself to an area of AI called natural language processing (NLP), which focuses on the construction of systems that process human languages, whether written or spoken. She was head of the group that developed the Automatic Grammar Checker for Portuguese in the 1990s, a pioneering project carried out in partnership with computer manufacturer Itautec-Philco and funded by FAPESP (see Pesquisa FAPESP issues 35, 46, and 58). Microsoft incorporated the tool into its Windows word processor for Brazilian consumers.
She is married to mathematician Wagner Vieira Leite Nunes, a retired professor at ICMC-USP, and her 30-year-old son Bruno is a psychologist. She also writes chronicles and short stories, which she posts online under the pseudonym Anelê Volpe.
Natural Language Processing (NLP)
Institution
Institute of Mathematical and Computer Sciences at the University of São Paulo (ICMC-USP)
Education
Bachelor’s degree in computer science from the Federal University of São Carlos (UFSCar, 1980), master’s degree in computer science from the University of São Paulo (USP, 1985), and PhD in computer science from the Pontifical Catholic University of Rio de Janeiro (PUC-RJ, 1991)
You are best known for the Portuguese grammar checker that your group created in the 1990s. What did you do after that?
Many things. My research is on a specific area of AI called natural language processing, which focuses on the construction of systems that process human languages, whether written or spoken. We call it natural language to differentiate it from other languages, such as programming code, graphics, mathematics, and others. These systems can review and correct text, translate from one language to another (machine translators), summarize or simplify a text, answer questions (chatbots), analyze coherence, and check the suitability of a text for a given purpose. Due to its complexity, the grammar checker project at ICMC-USP resulted in the creation of the Interinstitutional Center for Computational Linguistics (NILC), where many computational linguists studied before joining universities across Brazil and worldwide.
Your team was responsible for creating important resources needed for processing Brazilian Portuguese. What were they?
Dictionaries, lexicons, programs, in short: linguistic and computational resources that were needed to process the Portuguese language and that did not exist beforehand. Our group was well-known and became a unifying force for people across Brazil. We have been involved in many partnerships, agreements, and projects. The result was that other teams of researchers were also strengthened. Today there are active NLP groups at practically every university and educational institution in Brazil. The Portuguese NLP community has been holding the PROPOR conference [International Conference on Computational Processing of Portuguese Language] since 1993. It takes place every two years, alternating between Brazil and Portugal, to present the results of academic and technological research and to bring together research groups in the field. It was most recently held in Galicia, Spain, in March, since many consider that Galician and Portuguese are varieties of the same language.
Have you worked with the Portuguese and the Spanish?
We have a lot of contact with the Portuguese—PROPOR is the main initiative. Relations with the Spanish take place through cooperative scientific projects and student exchanges. In the late 1990s, we worked on a translation project with the University of Alicante. That is a topic we have been working on for a long time. Around the same time, we joined a major multilingual machine translation project at the United Nations University with representatives from each UN language. Unlike nowadays, translation systems at the time were made for each language pair. This project, which aimed to make use of an intermediate logical language called UNL [Universal Networking Language], provided a UNL decoder and encoder for each natural language. Each team was thus tasked with developing these two systems for their language. Once that was done for every language, anyone could translate text files from another language into their own and vice versa. At NILC, we are responsible for decoding and encoding the Portuguese language.
Our group at USP was well-known and became a unifying force for people across Brazil
Did it work?
No. For two reasons. One of them is technical: for this intermediate language to function well, it would have to be universal, capable of handling the meanings of all languages. Because it needed to be material in nature, they chose English. The English words were supposed to be capable of representing all concepts in all languages univocally, and that was not the case. The other limitation was political. Every team—and there were dozens—defended their own point of view. No agreement could ever be reached. It did not work out, but from an NLP perspective, we learned a lot. Several master’s and PhD projects were done at NILC with machine translation resources. A short time later, machine learning technology [programs that learn from data] took over in this area, leading to far superior automatic translators like Google Translate.
Besides the grammar checker, has your group released any other commercial products?
Not directly, no. Despite being founded with the purpose of creating a product, this is no longer the objective of NILC. For us, the most important thing is advances in knowledge. The center is a research and development center for basic resources for processing Portuguese, such as syntactic analyzers, semantics, and dictionaries, which are the basis of linguistic knowledge. We give unrestricted access to everything we create. Applications developed in academia normally serve as prototypes to demonstrate proposed ideas. Commercial products arise from partnerships with companies, as was the case with our grammar checker, which we developed together with Itautec. Nowadays, it is more common for companies to invest early in academic groups and then for their internal teams to finish the work started in academia. Some of the projects we are working on or have completed include prototypes for automatic summarizers [tools that produce summaries of news or documents], systems that simplify texts, adapting the language for children or adults with limited literacy skills, tools that detect fake news, assist with scientific writing, and transcribe audio, and more.
To what do you attribute the recent developments in AI?
Firstly, to the advancement of computing power. Modern computers are much faster and can be joined together to increase their power. Secondly, cloud computing, which increases computing capacity for many more people. This high processing capacity has allowed old artificial intelligence models—artificial neural networks—to show their true potential. Artificial neural networks [ANNs] are mathematical processing models inspired by the human brain. They are formed of processing layers that allow the program to learn a given concept based on many examples of that concept. The amount of data provided to the ANN to learn from is critical to the final result, as is the structure of the network. These models have been around for a long time, but they require a lot of data and a lot of computing power, which was not available in the past. Deep neural networks [deep learning] also revolutionized the field, with really significant results.
Does everyone have access to these technologies?
I will talk about NLP specifically. Large language models [LLMs] trained on huge amounts of data and parameters represent the language in which they were trained. They are thus able to generate text of quality comparable to that of humans—sometimes even superior. While it is possible to converse naturally with an LLM-based chatbot like ChatGPT, it does not actually understand the language. The “magic” occurs due to a complex and layered system of numerical weights. The fact is that currently, almost all NLP and AI problems can be solved with LLMs, which far outperform previous systems. But these models are very expensive to build because they require a lot of data, a lot of people to process it, very powerful servers, and a lot of electricity. Only major multinational technology companies—what we call “big tech”—have this capacity. To access these models, you need to pay for the best version or use a more limited one. You can even rent cloud servers to train your own model, but that is very expensive. The result is that the whole world is at the mercy of those who possess this technology. To a large extent, we have lost control over what these systems do. Even classic linguistic information such as spelling, grammar, word meanings, etc., is no longer really necessary. It is a major breaking point for NLP. We are one step away from throwing away everything we have done so far, to instead rely solely on LLMs going forward.
Is there a lack of criticism regarding the use of AI?
Yes. The people who develop AI need to make an ethical commitment, but I am not sure that they are aware of the scope of their decisions. Are they choosing data wisely, anticipating the impact of their decisions? I do not think so. Big tech companies hire lots of people from developing nations to collect data and train their software. The people working on this data are disconnected from the objectives of what they are doing. On the academic side, I still do not see a desire to change our behavior to draw attention to this. It is our duty to ask these questions. Just as we must raise awareness of how AI can help solve many of society’s problems—in medicine, agriculture, public security, and well-being—we also have an obligation to warn people about its limits and potential risks. If we do not have all the answers, if we do not know how to mitigate these problems, we can at least warn society so that the appropriate measures can be taken. This is what we are seeing several government agencies do around the world.
The people developing AI systems need to make an ethical commitment
Does this explain the need to regulate the use of AI?
Exactly. We have to establish regulations, as is being discussed in Brazil and other countries (see Pesquisa FAPESP issue n° 331), because generative AI models, including ChatGPT, have flaws. The biggest is that the algorithms are not explainable. Not even those who created the software really know what happens inside, because there are millions of combinations of numbers and functions. We also do not know how errors arise and who caused them. If something goes wrong, who should be held responsible: the machine, the programmer, or whoever fed the machine with its underlying data? And how can these errors be corrected? Nobody knows, because the programs are not explainable. This is why people say that generative AI is not transparent. The companies behind these programs need to tell their customers exactly how they work, what results can be expected, and in what situations they might go wrong. But that does not happen. Many people in the AI community are now researching ways to mitigate these shortcomings.
Because of the way the programs are built?
Yes. Because of the deep neural network model, language models, etc. We need rules, because these systems present a massive risk of generating unpredictable and potentially harmful results. The problem is that there are already people using these technologies without any rules to define how to deal with potentially unwanted consequences. Training language models with data from social media, for example. Due to the nature of social media, the linguistic material used to train the model may be contaminated with racism, xenophobia, homophobia, and other unwanted values, the propagation of which is harmful. We need to be very careful. But we are not being careful at all.
What are the biggest challenges currently faced by NLP?
The sudden success of LLMs has paradoxically become a challenge for the field. How can we adopt such an opaque and uncontrollable model as a solution to so many problems? But there are other obstacles too. One of the biggest challenges is how to computationally handle the semantics of natural language—how to create a system capable of choosing the correct meaning expressed by a sentence or text. Computer programs do a great job of understanding the shape of a language. We can teach them what a word is and how to form grammatically correct sentences, because we rationally express morphology and grammar. But the meaning is more difficult, because there are no formal rules to define the meaning, the significance of our words. It can seem like LLMs understand meaning, but that is only an illusion. Large language models know trillions of words in their context, as part of sentences, which allows them to “learn” that language. If you ask: “It’s cold today, what do you think I should wear?” the system will respond something like: “I suggest you wear a wool sweater.” The impression is that the machine understood the question.
But it did not understand?
It is an illusion. The machine has been trained on so much information that it simply provides the most likely sentence that should appear after the sequence of words with which it is presented. Most internet users ask a question. The software has no idea that it is a question—it only deals with sequences of words. It is us who give meaning to the response, deeming it an appropriate answer to the question. We also do not know how our own understanding is processed, how we give meaning to words and things. How can we expect to build a machine capable of doing something that we ourselves do not even know how we do? And there are other problems. Like every AI system, NLP results are not exact, they are always approximations of ideal solutions. Machine translation is very good, but it is not perfect; an automatic summarizer can be very useful, but it is not perfect. We need to identify the ethical problems and errors in these programs while we are developing them, not afterwards. If an algorithm presents any situation of risk, we have to prevent it and correct it. Instead of trying to find the optimal solution at any cost, why can we not accept the best solution that avoids risks and harm?
How can these problems be minimized?
What is within our reach is to demonstrate the knowledge we have and clarify how these things happen. That is one of the aims of the Brazilian NLP group that I am a part of, which already has 211 participants. In 2023 we released a book called Processamento de linguagem natural: Conceitos, técnicas e aplicações em português [Natural language processing: Concepts, techniques, and applications in Portuguese]. It is available for free online. It is more comprehensive than any existing Portuguese textbook on NLP. Helena Caselli of UFSCar and I organized the book, which has more than 60 authors. We want to keep it updated and show people what is going on behind Portuguese language NLP systems. An expanded, updated third edition of the book will be released in November.
Punched cards and floppy disks used in the early days of computingPersonal archive
Are there other initiatives?
Since we founded NILC, we have been working on more resources for Portuguese, like a big corpus to train LLMs, while ensuring it is comprehensive and varied enough to be representative of the language we want to model. There are other research groups doing the same, such as the Artificial Intelligence Center, which I am a member of and which is funded by FAPESP and IBM. But this all demands time, money, and human resources. At the end of 2022, the world was stunned by a startup called OpenAI, which launched ChatGPT, a chatbot that uses natural language models, including in Portuguese. I see this great desire to produce systems quickly, because the new models allow you to do so. These systems can be used in schools, by children, even though they are very poorly tested. ChatGPT is an example. When it first appeared, everyone thought it was incredible, but a few days later they realized that it was producing a lot of false information.
What led you to become interested in programming?
I’m from Sertãozinho, a town in São Paulo State, and I came to São Carlos in the mid-1970s to take a library science course. I was already fascinated by being around books, although I was not actually aware of it at the time. But the course only lasted six months. I needed a bigger challenge. Since I was already here, I looked into courses at both universities: UFSCar and USP. I chose to take a computer science course at UFSCar, which at the time was seen as the profession of the future, although I had no idea what it was. In fact, in the first year of computing—in 1977—we had no contact with computers. We had just one minicomputer, mini only in name, because it was a huge machine. It was installed in a large, air-conditioned room, with one person receiving the cards and punched tapes that were used to encode programs.
What did you study for your master’s degree and PhD?
I did my master’s degree at USP on a topic known as algorithm analysis, supervised by Professor Maria Carolina Monard [1941–2022]. I studied which algorithm was best at searching for words in text files, long before the internet, Google, and other search tools existed. I did my PhD at PUC Rio under the supervision of computer scientist Tarcísio Pequeno, who at the time worked with the theory of computation. It was there that I met Professor Clarice Sieckenius de Souza and Professor Donia Scott, a Jamaican from the University of Brighton, England, who was there on a sabbatical year. I started working on NLP with them. I finished my PhD, returned to USP in São Carlos, where I was already working as a professor, and a short time later, we started the grammar checker project. From then on, I immersed myself in NLP and began working exclusively in this field.
How important is Monard to Brazilian AI?
She pioneered AI research in the country. As a professor at ICMC-USP’s Department of Computer Science, she has supervised many researchers since the 1980s. She has always carried out high-quality research and laid much of the groundwork in machine learning, which is behind the major recent advances in artificial intelligence. In predominantly male environments, such as the exact sciences and the technology sector, Monard was a role model for many young women. Unfortunately, the number of women choosing to study computing has fallen significantly in recent decades. When I was a student at UFSCar, in only the third computer science class at the university, the number of men and women was more or less equal. As society became more computerized and computing professions became more popular—not that that is necessarily the cause—the number of women in the field fell significantly. I have noticed, however, that there are now several movements trying to include women in the exact sciences in general, and in computing in particular. I believe these interests rise and fall over time, depending on factors beyond a person’s aptitude. Raising awareness of the field, breaking down gender prejudices, and having women set the example by occupying prominent positions are all effective ways of reversing this situation.
In the past, the number of men and women in computing was equal. Now, the number of women has fallen significantly
Why did you retire in 2013, before the age of 60?
I retired because I had worked for the number of years needed to do so. I thought I had already contributed enough. I was head of the undergraduate and postgraduate computer science programs and I participated in every committee possible and imaginable. I decided that they no longer needed me, and that in the classroom, students would benefit more from younger professors who had more recently completed their studies, since knowledge in technology is constantly evolving. Our research group at NILC was very well managed by Professor Thiago Pardo, who was a doctoral student of mine. They did not need me there anymore either.
So you officially retired, but you continued working, right?
Yes. I am still at USP as a senior professor. I teach the occasional class and follow NILC’s work. One NILC project is Poetisa, through which several groups are putting together a corpus of written and spoken Brazilian Portuguese. Among other purposes, these texts can be used to train algorithms that perform syntactic analysis [parsing], an important task for most NLP applications. There are several parsers in Portuguese, but we have room to improve. This is a job that will take many years. Another project is looking at speech processing, led by ICMC-USP professor Sandra Aluísio. The aim is to create transcriptions of spoken language from audio, with representative samples from across Brazil, for use in applications that will use Portuguese audio.
When did your writing career begin?
It has long been a desire of mine, but maybe in the past I did not recognize it. Now I have realized that if I had not been a teacher, I would have loved to have been a writer. I started writing in 2005, before I retired. I wrote some chronicles, I liked them, and so I decided to continue. In 2007, I discovered an app called 700 Words. It encourages you to write 700 words a day for 30 days. I took on the challenge and wrote a column every day for a month. It was a lot of fun. Now I have my own website, which I created in the early days of the pandemic to publish everything I had already written. I didn’t want to publish it in a book because a lot of it was very personal. I would have felt embarrassed. But on a website, under the pseudonym Anelê Volpe, I feel more comfortable. I did publish a book about my family history, however, to give to my relatives.
