Research Computing Sciences

We Can Help Language Technologies Understand What We Want

, by Andrea Costa
Debora Nozza wins an ERC Starting Grant for a project that will change our relationship with advanced language based technologies

One big problem with digital technologies is that, despite collecting huge amounts of information about people, their output is largely standardized. More often than not, similar queries by different people are bound to produce similar results.

Debora Nozza of Bocconi's Department of Computing Sciences has won a €1.5m ERC Starting Grant for her project, called PERSONAE (Personalized and Subjective approaches to Natural Language Processing), whose goal is to make language technology (LT) accessible and valuable to everyone.

No two people are alike. Sensitivities can and do change across space and time. It is therefore unrealistic, and ultimately unfair, to assume that everybody needs the same content moderation, the same suggestions, the same responses. The underlying principle of this project is that everyone should be able to decide what information to share with any digital interface (social networks and virtual assistants are but two common examples) and should have the right to know how this information has been used and to what purpose.

Everyone agrees that LT can prove very useful in countless situations. Debora Nozza wants to make LT applications more controllable by users so that its utility can be further boosted. The advantage would be twofold: a user would receive more useful output and would be able to learn how to improve responses. On the other hand, a provider would be considered more trustworthy and reliable if results were increasingly transparent and tailor-made on the basis of whatever information a user has freely decided to share. The resulting environment would potentially have real advantages in terms of better quality of life.

This entails a rethink of how LT algorithms work, which is part of the scope of this project. It proposes transforming the present research aims, encouraging more researchers and common users to participate and providing an entirely new perspective.

"Language processing models solving subjective tasks should be designed by individual users for their specific needs. We should build models that consider individual perspectives," says Debora Nozza. "Users will be able to hide content that they don't want or choose to receive content that they personally find humorous (funny cats but not funny dogs). They will also be able to influence how these preferences are accounted for over time, receiving more entertaining content after a long day at work, for example. PERSONAE will change how LT models are designed, making individuals active participants in the process instead of mere recipients. This will allow for a much more tailored, effective approach to language-processing model design, resulting in better models overall."

ChatGPT, Listen to Me!

Watch video