Horacio Saggion, Full Professor of Computer Science and Artificial Intelligence at UPF
Head of the Large-Scale Text Understanding Systems (LaSTUS) lab of the Natural Language Processing (TALN) research group in the Department of Information and Communication Technologies (DTIC) at UPF

In recent months, news portals and social media have been brimming with news about ChatGPT, a system trained to generate sequences of words (or source code) from a context or text fed to it as a prompt. 

The scientific paper describing this technology makes it clear that this generalist text-generating system significantly outperforms its predecessors when it comes to various tasks requiring the use of language (summarizing texts, answering questions, translating texts, etc.). This is due, first, to the computational architecture used (Google’s deep-learning Transformers architecture, widely adopted by the scientific community) and, second, to the large number of parameters compiling the knowledge about language that the system has acquired and which it has been trained on with millions of texts. Unlike other systems, GPT-3, the language model on which ChatGPT is based, is a generalist model and not specialized in specific tasks. 

Although they cannot be considered formal tests of the model’s capabilities, a handful of interactions with the ChatGPT application is enough to impress anyone. The generated texts seem coherent and grammatically correct. The system also excels at proposing solutions to simple programming problems. Indeed, it goes even further, as there seems to be no topic on which ChatGPT has nothing to say. For example, if you ask it about issues such as homosexuality in sport, it is able to give reliable information without discriminatory over or undertones, as well as paraphrase existing regulations on discrimination in sport. It is likewise capable of admitting some mistakes when questioned. 

This system is a true breakthrough in the field of automatic text generation, which, for years, was considered the poor relative of natural language processing, but has now proved to be the richest

This system is a true breakthrough in the field of automatic text generation, which, for years, was considered the poor relative of natural language processing, but has now proved to be the richest. In my view, this is not a case of much ado about nothing, but of much ado about a tool with enormous potential for many sectors of society, such as the hundreds of startups that are leveraging it, as seen at the recent Barcelona Mobile World Congress. 

However, ChatGPT is not the end of the road for natural language processing research. Many areas remain to be explored, in particular to provide technological solutions for those languages for which there are not enough data to develop natural language processing systems (sign languages, minority languages, etc.) or for applications requiring the production of reliable and verified information (scientific documents, medical reports, highly complex legal documents, etc.). Human-machine collaboration and care for ethical considerations will be fundamental in the coming years. 

Just as we have grown accustomed to electronic calculators, industrial robots, or medical technology, to name just a few examples, we will see these advanced conversational systems and marvel at their capabilities, whilst also criticizing their mistakes, even as we get used to them, too.