The Chair of Artificial Intelligence and Language Technology has an ambitious program to strengthen leadership in this technology and place the country at the technological forefront. To achieve this, it is supported by two pillars: On the one hand, scientific excellence at the HiTZ research center, Basque Center for Language Technology of the UPV/EHU, in collaboration with the Faculty of Informatics of the UPV/EHU. On the other hand, the company Avature, whose transparent and intentional approach regarding the development of artificial intelligence makes their platform to be chosen by many international companies and government entities to safely deploy their technology in talent-related tasks. The main objective is to reinforce the research leadership with measures to increase the training of experts, the synergy with the industry, the social awareness of the real risks and research into more eficient and unbiased language models. There are four concerted lines of action planned for the next four years: 1) Increase the number of experts through undergraduate, postgraduate and continuous training ofer. 2) Technological transfer through a novel collaboration mechanism. 3) Foster research into generative Artificial Intelligence, improving aspects such as energy eficiency and biases. 4) Improve the projection of research to society through a dissemination plan with special emphasis on generative AI, since a society better informed about the capabilities of AI will be able to adopt and enhance the benefits, and in turn identify and mitigate the new challenges it entails.
• Increase the number of experts through undergraduate and postgraduate studies, as well as the continuous training of technicians, managers and into their systems. Each of these models can be used on social agents in this technology. many tasks, such as search engines, chatbots, summary • Transfer between the chair and companies creation, translation or computer program generation. through a novel collaboration mechanism that seeks the synergy between them, detecting the Motivation of the Chair most promising areas. • Foster research into generative Artificial Intel- The NLP community is contributing to the emergence ligence, improving aspects such as energy efi- of disruptive new DL techniques and tools that are revciency and biases. olutionizing the approach to language technology LT • Improve the projection of research to society tasks. The typical way to implement NLP solutions has through a dissemination plan with special em- moved from a pipeline based methodology, to architecphasis on generative AI, since a society better tures based on complex neural networks trained with informed about the capabilities of AI will be able vast amounts of data. This rapid progress in NLP has to adopt and enhance the benefits, and in turn been possible because of the confluence of four diferent identify and mitigate the new challenges it en- research trends: 1) mature deep learning technology, 2) tails. large amounts of data (and for TL, large and diverse multilingual textual data), 3) increase in high-performance
Before looking at each of the lines of action, we present computing power in the form of Graphic Processing Units the motivation for this chair, and the needs detected. (GPUs), and 4) application of simple but efective selflearning and transfer learning approaches using Trans2. Motivation formers [1, 2, 3, 4].
Thanks to these recent advancements, the NLP comThe ability to communicate through natural language is munity is currently engaged in a paradigm shift with a key aspect of human communication and is essential the production and exploitation of large, pre-trained to efectively transmit information. Natural language transformer-based language models [5, 6]. As a result, processing (NLP) and language technology (TL) are fields various IT corporations have started deploying large prewithin Artificial Intelligence (AI) that focus on the devel- trained neural language models in production. For inopment of algorithms and systems that can understand stance, Google and Microsoft have integrated them in and interpret human language. NLP is an important area their search engines. Compared to the previous state of of AI with numerous applications, particularly in the con- the art, the results are so good that systems are claimed text of today’s digital transformation. Language is and to obtain human-level performance in laboratory benchmust be at the center of our eforts to develop AI, and marks when testing some dificult language understandvice versa. In fact, currently TL is arguably the most inno- ing tasks. vative field of AI with rapidly growing economic impact. Due to the great results obtained, nowadays there is Both AI and NLP are undoubtedly expanding fields that a tendency to build increasingly larger models. For ingenerate an immense volume of business and qualified stance, GPT-3 contains 175 billions of parameters and was jobs. The global NLP market is estimated to be worth trained on 570 gigabytes of text, with a cost estimated $13.5 billion in 2021 and is believed to reach an expected between ten and twenty million USD. In comparison, its value of $91 billion by 2030, growing at a CAGR of 27% predecessor GPT-2 was 100 times smaller, with 1.5 bilduring the forecast period (2022-2030). These figures indi- lions of parameters [7]. This upscaling leads to surprising cate that the return on investment will be enormous. For behavior: GPT-3 is able to resolve tasks for which they example, OpenAI, owner of ChatGPT, estimates one bil- have not been previously trained, just by providing them lion dollars in revenue by 2024. It is therefore necessary with very few training examples. This behavior was not to guarantee competitiveness in this area, through tech- observed in the much smaller GPT-2 model. There are nological surveillance, training, dissemination, transfer works that indicate that this behavior depends directly and cutting-edge research, guaranteeing strengthening on the scale [8], so that, from a specific size, models show the entire TL ecosystem of the country. emerging abilities that allow them to undertake tasks for
In recent years the emergence of new and powerful which they have not been taught.
Deep Learning (DL) techniques are revolutionizing NLP. Large language models obtain remarkable results but Neural machine translation systems or language models are extremely costly to train and develop, both finansuch as ChatGPT or Bard allow developing applications cially, due to the cost of hardware and electricity or cloud that were unthinkable just a few years ago. As a result, computing time, and environmentally, due to the carbon leading technology companies such as Google, Facebook, footprint required to fuel modern servers with multiple Microsoft or Amazon have integrated large neural lan- Graphics Processing Unit (GPU) hardware. This also guage models pre-trained with huge amounts of data means that only a limited number of organisations with abundant resources in terms of funding, computing ca- niques they study, and they could greatly benefit by colpabilities, NLP experts and data can currently aford to laborating with Chair’s entities when developing their develop and deploy such models, thus aggravating the final degree or master’s thesis. Likewise, the lack of comlack of technology sovereignty in most linguistic com- putational resources available to students limit the scope munities. of their work. On the other hand, technical staf in com
There are also worrying shortcomings in the text cor- panies, research/technological centers and administrapora used to train these anglo-centric models, such as tions are demanding continuous education programmes the predominance of harmful stereotypes or the lack of to widen their knowledge base and enhance their technirepresentation of less-resource languages (which in fact cal skills in these latest advances. This is applicable not represent the vast majority of languages, and include only to software engineers, but also to technicians from Catalan, Basque and Galician). Language models are other areas who now see how AI&TL are causing disrupknown to be biased in several respects, and they also lack tion in their domains. Likewise, the managing staf in explainability, that is, the decisions made by these models companies, technology/research centers, administrations cannot be scrutinized, due to their black box nature. are often unaware of the opportunities and threats that
On the other hand, we must be aware of the risks of the AI&TL technologies pose. massive use of AI&TL in society. For example, this tech- It is also necessary to devise a good dissemination nology usually works much better in certain languages strategy to raise the conscience about AI&TL in society, with large number of speakers, while the vast majority both their possible benefits but also its harms. New techof the languages run the risk of being excluded from nologies tend to generate opinions that are not always their benefits, thus losing attractiveness and generating well-founded, both on the exaggeratedly optimistic side a worrying sociolinguistic impact. On the other hand, (AI&TL like ChatGPT is an intelligent being like humans), biases may have harmful efects on society. In another and on the catastrophic side (AI&TL is so powerful that example, neural translation and text generation systems it will leave screenwriters out of work). These exaggermay produce worrying changes in the information flow, ations can alter the perception that general public has such as the spread of fake news or plagiarism. These regarding subtle efects, such as biases in automatic transrisks are also worry for companies that use TL, as their lation or the real impact of text generation systems in products are sensitive to these risks, which can lead to classrooms, which are therefore not known to their full suboptimal results and, in some cases, serious reputation extent. problems. Regarding transfer as well as industry and administrations needs, there is a general unawareness of the Needs possibilities of AI&TL, and agents and practitioners need a framework where they can share their needs and capabilities, and hence be able to design processes and applications that do not exist today.
Finally, society requires research that takes the needs of companies and administrations into consideration, incorporating realistic scenarios and data, or building systems that can be deployed in environments where the computational requirements are constrained. The Chair will constantly monitor and detect new needs, and adapt its actions accordingly.
The necessity for further exploration and research in the ifeld of AI&TL is evident. Modern AI&TL is mostly based on deep neural networks, a key-enabling technology that has shown enormous potential in the productive and social fabric. In this context, it is paramount for the university and society in general to react, and to avoid being mere passive spectators of these disruptive changes.
The University of the Basque Country already works on AI&TL within its educational, transfer, and research activities. However, the Chair aims to go a further step and become an institutional unit researching on AI&TL that systematically helps generating an explicit reference point inside the university itself, therefore contributing to visualize, promote, energize and join all eforts towards this direction, both in the university and in society.
Besides, the Chair should also undertake dissemination and dynamization tasks in society. It has to serve as a connection agent that facilitates a knowledge flow from academy towards society, and in turn feed by its needs and concerns.
From a social point of view, there are urgent needs regarding education. Bachelor’s and master’s students are often unaware of the specific applications of the techSocial Relevance A society that is properly informed about the capabilities of AI&TL will be able to adopt and enhance its benefits while identifying and mitigating the new challenges they may pose. Administration directors who are properly trained and informed about AI&TL will be able to make efective decisions that improve society in general. In turn, a forum where companies developing AI&TL, consumer companies and university researchers share their potentials and needs will allow for more efective transfer, research oriented to society, and ultimately a synergy and mutual benefit.
annual event to promote the exploitation of the results of
the chair and AI< technology. (
Regarding the training of undergraduate and master do today. The latest advances obtained in AI< by the
students, the Chair will serve to reinforce the training participants of the chair will be presented. Several
colof students through their final degree projects. To this laborative projects will be chosen by the participants of
end, Final Degree Projects and Master’s Degree Projects the chair, ofering the servers owned by the chair when
will be defined in the subject of the Chair in bachelor’s possible.
and master’s degrees related to AI<. The
participation of the entities of the Chair will be encouraged in
the definition of these projects and in their co-direction. 5. Research
Likewise, internships will be defined and carried out in
the companies and institutions associated to the Chair, The following activities will be encouraged: (
Regarding continuous training, it is planned to ofer, applied aspect, in the sense of focusing on real problems
together with the entities of the Chair, an own title of that concern the business entities part of the chair. (
The chair will apply the following actions for the
exploitation of research, training and dissemination results with
measures aimed at increasing the impact of the results
on the productive/scientific, social and economic fabric.
(
ISG Intelligent Systems Group. The Intelligent Systems Group (ISG)5 focuses its research in the areas of machine learning, combinatorial optimization and high The chair is closely linked to the HiTZ Basque Center performance computing, where it has an extensive refor Language Technology1 and the Computer Science search experience. ISG is one of the UPV/EHU groups Faculty, both in the University of the Basque Country with a consolidated category "A" group in the latest call UPV/EHU. It comprises the research groups below. The for research groups of the Basque Government. Its voother pillar of the chair is Avature, also described below. cation for high-quality basic research is reflected in the more than 200 publications in international journals with impact factor, and in the 16 doctoral theses defended during the last 5 years.
Aholab Signal Processing Laboratory. Aholab2 is the short name of the Signal Processing Laboratory of the University of the Basque Country (UPV/EHU). The laboratory is located in Bilbao. We are a university re- RSAIT Robotika eta Sistema Adimendunen Taldea. search team and focus our research in the areas of Text The Robotics and Autonomous Systems research group to Speech Conversion, Speech and Speaker Recognition, develops its activity at the Computer Science Faculty of and Speech Processing in general. Since 2005 we are a rec- the UPV/EHU since 2005. Nowadays, its research eforts ognized research group of the Basque Research Network. focus in two main areas: the development and applicaThe laboratory is part of the Basque Center for Language tion of machine learning techniques, mainly for activity Technology (HiTZ) and the Department of Communica- recognition in videos; and the semantic based co-speech tions Engineering of the Faculty of Engineering of Bilbao gesture generation in social robots. It is a consolidated (EIB). A group according to the Basque Government and has a long participatory tradition in cultural events, disseminating the state of development of both, AI and robotics. 1https://hitz.ehu.eus 2https://aholab.ehu.eus/ 3htpps://ixa.ehu.eus 4https://www.aldapa.eus
versity of the Basque Country (UPV/EHU) that works in all areas of Natural Language Processing. Ixa is a Avature. Avature6 is a highly configurable softwaremultidisciplinary group with more than 25 years of ex- as-a-service (SaaS) platform for enterprise Human Capiperience, comprising computer scientists, linguists and tal Management (HCM), focusing on talent acquisition other disciplines. The group is based on the Computer and management. It is the leading provider of CRM Science Faculty in San Sebastian and the Languages and and ATS technology for human resources worldwide. Computer Systems department, but many members be- Founded by Dimitri Boylan, co-founder and former CEO long to other faculties and departments of the UPV/EHU. of HotJobs.com, Avature brings consumer-driven inThe group is part of the Basque Center for Language ternet technology and digital innovation to human reTechnology (HiTZ). sources departments. Avature solutions are used by 110 Fortune 500 companies in more than 164 countries and in 32 languages. It delivers its services from its private cloud located in data centers in the US, Europe and Asia.
Avature has ofices in Buenos Aires, London, Madrid, Melbourne, Munich, New York, Paris, Shenzhen and Virginia.
and PArallelism) is a research group from the University of the Basque Country (UPV/EHU) mainly based on the Computer Architecture and Technology department of the Computer Science Faculty but it also includes staf from other engineering schools in the UPV/EHU. The group has more than 25 years of experience using, proposing and adapting machine learning algorithms to solve real world problems, having currently two main focuses: artificial intelligence and physiological computing in health, specifically in diagnosis and prevention of nervous system diseases (Parkinson’s, Alzheimer’s, stress, epilepsy), and, fair and explainable machine learning models.
This chair is funded by Avature and the Ministerio para la Transformación Digital y de la Función Pública in the context of the Estrategia Nacional de Inteligencia Artificial (TSI-100923-2023-1) and the European Union NextGenerationEU/PRTR.
5http://www.sc.ehu.es/ccwbayes/ 6www.avature.net