AI: A Lifeline for Minority Languages?

Pilar Manchón (Seville, 1972) is a philologist and computer scientist, and has been the director of artificial intelligence research strategy at Google since 2019. She arrived in Silicon Valley twelve years ago and, since then, her career in the world of computational linguistics has only grown. She has also been a speaker at the Talent Arena, the parallel congress to Mobile that was held in Montjuïc. Manchón's career has not been linear and in a conversationCompanies He explains how he got to where he is now and what the world of AI can contribute to language.

"When I started university, I went into physics, but I had a sports accident and I was left out of the loop," he says. For Manchón, this accident marked a turning point, and he decided to go to England to study English. "I had always liked languages, and so I decided I wouldn't study physics, I would study philology." When he reached his final year, he went to Ghent (Belgium) on a scholarship to study translation and interpreting. "There I began to see language technologies that seemed very interesting to me, and I began to discover a very rudimentary tool that existed at that time with which translators and interpreters could work partially with technology," he explains.

"I started to follow the thread, to inform myself, and I began to study computational linguistics. In the end, I ended up applying for a scholarship to do a master's degree in cognitive science and natural language at the University of Edinburgh, where I learned to program," says Manchón. "My master's thesis was on language technologies, basically on how to make speech synthesis much more natural. And then I enrolled in a European doctorate program, where my thesis was on intelligence in multimodal dialogue systems," explains the expert. Then, through a Fulbright scholarship, she went to Stanford University. When she finished, she returned to Seville to complete her doctorate, and realized that the technology her doctoral research group was working on was "far superior" to the technology she had used in America.

Given this, Manchón and her thesis professor ended up launching a company using this technology. "Ten years later, we managed to take it internationally and compete with other companies in the sector. It all culminated when we met Intel; they really liked us, so much so that they wanted to invest in us. At that time, Siri (Apple's voice assistant) had been released just a few years earlier. Intel was about to launch its own phone, a company with the goal of turning it into Intel's Siri," he explains. Ultimately, Intel didn't launch its phone, but this is the story of how they were acquired, and how Manchón ended up in Silicon Valley. That was 12 years ago. "I worked for Intel for several years, then for Amazon. Later, I spent a year and a half with Roku, a streaming and television device company. And I've been with Google for five and a half years now."

Trust in the Machine

At Google, Manchón is primarily dedicated to applied research in the field of computational language, which serves primarily to generate trust in machines. "I'm an expert in communication with machines, and I've always been very interested in all topics related to trust. How do we develop trust? From a cognitive perspective, with intelligent entities that can speak to us. I've dedicated myself to developing projects in which we've seen how trust evolves between a human person and an intelligent entity, and which ones. This translates to building voice interfaces, more natural interfaces, with different product areas within Google. "We've also delved into topics of optimization in the impact processes of large models. In translation, for example, things are being done with generative artificial intelligence. And, recently, more topics of factuality, which is making sure the models tell the truth, or also align with human values," he points out.

AI for minority languages

As an expert in computational linguistics, Manchón argues that AI can be a huge lifeline for minority languages. "One of the great advantages of these technologies is that once you have a pre-trained model to translate, for example, from one language to another, you have a very strong foundation for the model to learn another language with very little data," she explains. "There are some languages for which we have a lot of resources, such as English, Spanish, or Chinese. But there are many other languages in the world, around 7,000, and many of them have very few resources compared to the majority languages. The advantage that this type of technology offers is that with few resources you can achieve more than those that exist in that language. And now, precisely with these techniques, with a pre-trained model, you can use these few resources to achieve a very high degree of performance.

Catalan has been one of the most requested languages by users in voice agents, such as Amazon's Alexa, and there is still no date for Alexa to speak Catalan. According to Manchón, this is a matter of market needs: "From Google's point of view, what we do is make all the languages we can as quickly as possible, but also in accordance with market and product priorities. "I understand that Alex must do something similar," he says. "When it affects you closely because it's your language, it's a priority for you, but from the perspective of a global market, perhaps not," he points out. "There are many solutions for this type of thing. There are many free resources. There are many free distribution models. If there's a lot of interest in a language, anyone—in the academic world, any university, any researcher, any public entity—can develop their own model and generate whatever language they want, if they want," he counters.