Mots/Machines #6
Science, AI, Creativity and Creation
Program:
- 8h45 Reception
- 9h00 Opening
- 9h10 Juliette Le Berrigot, Master student writing/translation UBO, AI translation of Pokémon names
Details
Presentation
Résumé:
I'll be presenting the results of my research (English Pokémon fanmade names translated by DeepL and by ChatGPT in French).
Biographie :
I'm a student in M2 Writing/Translation at the UBO in Brest.
- 9h30 Jaap Kamps, faculty of humanities of University of Amsterdam - The Role of AI and Creativity in Scientific Inquiry
Presentation
Details
Résumé:
Creativity and creation are mostly viewed as the opposite of scientific rigor in science. The context of scientific justification is characterized by very strict rules, guided by formal logic and rigorous methodology. The context of discovery, however, has far more freedom: it is even best left to psychology according to eminent philosopher Karl Popper. Whereas logic dominated classical Artificial Intelligence (AI), recent revolutionary progress is with AI models that excel at generation, creation, and creativity. What is the roles of these AI models in scientific inquiry? What strength and what issues do these models have in this context, as they are known to “hallucinate” and exhibit confirmation bias? Whereas earlier models focused on the context of scientific justification, can these models play a role in the context of scientific discovery? And how does this change the role and task division between the human researcher and the AI models assistant?
Biographie :
Jaap Kamps obtained a PhD in “logical” artificial intelligence at the University of Amsterdam. He is co-founder of the University of Amsterdam’s Information Retrieval (IR) group, and its Natural Language Processing (NLP) group. He has worked on a broad range of topics covering user-centered to system-centered IR, including pioneering work on structured document retrieval, and on neural ranking. He is has work on many areas of NLP, including pioneering work on sentiment analysis, and on language modeling and text generation. Current interests are in “AI for social good” by working on novel access tools for cultural heritage and political data, and by developing explainable and interpretable neural models for search and recommendation, and ways to open up scientific articles and government information to laypersons, citizens, and journalists.
He has published over 450 papers in all major conferences and journals, which can be found at http://e.humanities.uva.nl/ ; https://scholar.google.com/citations?user=bWlQ2uEAAAAJ; http://dl.acm.org/author_page.cfm?id=81100034443 ; or other repositories.
- 10h20 Coffee Break
- 10h50 Christophe Servan - Considerations after the impact of Large Language Models and the arrival of Large Agent Models: towards a revival of multi-agent systems in AI?
Presentation
- 11h40 Helen McCombie, University Translation Bureau - My digital songwriting partner
Presentation
Details
Abstract:
I explore some ways digital tools can create, assist and inspire song lyric writing by presenting some examples of language treatment tools specific to this activity and some experiments using generically-trained LLMs.
Many song styles pose constraints on the lyric text needed, particularly in terms of word choice, including the selection of rhymes suitable for the subject context, and the adherence to stress patterns dictated by the music or existing lines of text.
While digital tools can assist with these aspects, wholesale writing of complete sets of lyrics by LLM-based tools creates a paradox. The generic nature of responses tends to produce lyrics lacking in specifics that would otherwise add relatability, plausibility and memorability. Using Chat GPT I tested queries that could channel LLM output into something less generic.
Songwriters welcome AI tools with varying enthusiasm. I make a round-up of some impressions gathered among an online songwriting community.
Finally, I take look at a few examples of interactive and commercial applications.
Bio :
Helen McCombie works at the university translation bureau where she specialises in scientific text revision. She is also an amateur musician.
- 12h00 Lunch
- 14h00 Adam Jatowt Estimating, University of Innsbruck, Austria - Temporal Validity of Text
Presentation
Details
Abstract:
It is important to learn whether information is still valid or not for various downstream applications including recommender systems, information retrieval, and user state tracking on microblogs and via chatbot conversations. It is also beneficial to deeply understand the story by tracking implicit information about the durations of protagonists' activities and involved events. However, this kind of inference is still difficult for machines as it usually requires temporal commonsense knowledge and reasoning. We propose and investigate a series of novel tasks related to temporal commonsense reasoning such as temporal validity estimation, temporal validity reassessment, and temporal validity change prediction of an input text given some follow-up context. In essence, these tasks require inference whether actions expressed in text are still ongoing or have been completed, hence whether the describing them content remains valid, or has rather become obsolete, either due to the elapsed time or based on the provision of additional context. Additionally, we also discuss several novel datasets that we have constructed for probing LLMs and NLP models in general when it comes to temporal validity estimation and reasoning.
Bio :
Adam Jatowt is a Full Professor at the Department of Computer Science of the University of Innsbruck, Austria. He also serves as a Deputy Head of the Digital Science Center and Deputy Head of the Research Center Digital Humanities at the University of Innsbruck. Adam received his Ph.D. degree in Information Science & Technology from the University of Tokyo in 2005, and afterwards he worked at Kyoto University for 14 years, first as an Assistant and later as an Associate Professor. His research interests lie in the intersection of natural language processing, information retrieval and artificial intelligence. Adam is on the editorial board of IP&M, JASIST, IJDL, and JIIS journals, as well as serves as a Senior PC member of SIGIR, WSDM, CIKM, ECIR, SIGIR-AP and JCDL conferences. He is a recipient of the Friedrich Wilhelm Bessel Research Award by the Humboldt Society and the Karlsruhe Institute of Technology’s (KIT) International Excellence Fellowship.
- 14h50 Riwal Lefort, Crédit Mutuel Arkéa - Artificial Intelligence in bancassurance
Presentation
Details
Abstract:
In this presentation, we address the subject of Artificial Intelligence (AI) in bancassurance.
After giving a vision of AI and a few general definitions, we present a large number of use cases specific to the bancassurance sector.
We will see that AI can occupy the entire information system, from the front office (customer relations) to the back office (data center management), via the middle offices (decision support).
We'll also see that all types of data are present: textual data (e-mails, news articles, etc.), images (invoice scans, account statements, etc.), bank transaction labels, etc.
Next, we'll look at the specifics and workflow of an AI project in bancassurance. Indeed, regulations impose strict constraints on the monitoring and explicability of AI models.
We conclude with a discussion on the adoption of generative AI: can it be used? What precautions need to be taken?
Bio :
After 10 years of academic research in Machine Learning for computer vision, bioinformatics or underwater acoustics, Riwal LEFORT was recruited in 2017 at Crédit Mutuel Arkéa (CMA) to develop Artificial Intelligence (AI) in the group. His work at CMA focuses on setting up and monitoring AI projects, but he also takes part in internal AI training courses and helps formalize AI projects (procedures and protocols).
- 15h40 Coffee Break
- 16h10 Jean-Charles MEUNIER Institut des Sciences Humaines, Université Polytechnique Hauts-de-France - Machine Translation, Multimodality and Puns: Limits and Prospects
Presentation
Details
Abstract:
As the focus of AI translation is primarily to transfer the sense rather than the sound, the translation of texts that include plays on language is a real challenge. This is made all the more difficult when puns rely on other modes, such as images in the case of subtitling. As Adrián Fuentes-Luque has shown in the case of films by the Marx Brothers, for example, the humour rests on the simultaneity of the image with the translated pun. These obstacles shall be explored through the case study of the short stop-motion animation Grocery Store Wars. The film uses the public’s knowledge of the famous Star Wars saga, in particular the opposition between a bright and a dark side, to denounce the use of genetically modified organisms and to promote the consumption of organic food. Machine translations by DeepL and ChatGPT shall be compared with human translations by students and by the presenter himself. This comparison shall be used not only to demonstrate the limits of machine translation, but also to suggest future developments.
Bio :
Jean-Charles Meunier teaches English language and culture, as well as translation studies, at the Université Polytechnique Hauts-de-France in Valenciennes. He has published several in-depth articles about Bob Dylan’s songs and has given talks on the topic at international conferences. His PhD thesis, entitled Multimodal Refractions of Bob Dylan in French Covers, explores Dylan's songs translated and performed in French over a time span of more than 50 years. In this study, he addresses issues related to metrics and musical adaptation, taking into account Dylan’s idiosyncrasies. He approaches the topic of song translation through the lens of multimodality, i.e. investigating the relationships between text, voice, music and sound and how these converge to create meaning. Great attention is also paid to historical and cultural contexts, in particular to the way culture specific references are transferred within or between modes.
- 17h00 Michel Delarche (online), Automatic poetry translation
Presentation
Details
Abstract:
This presentation deals with the current limitations of statistical machine translation systems for dealing with poetry, and in particular with the constraints of versification.
A comparison of a small example (a Shakespeare quatrain) with the output of human translators shows that current machine translation systems suffer from a double handicap: their limited ability to take context into account, and the unsuitability of their statistical databases for processing poetic language.
Research carried out over the past fifteen years to improve the performance of these systems in the face of the multiplicity of constraints to be satisfied is then reviewed, and three ways of making combinatorial exploration more locally flexible are proposed, based on an analysis of the strategies employed by human translators.
In conclusion, a method for targeted enrichment of training corpora is proposed.
Bio :
Michel Delarche born 1955
ENSIMAG engineer (1977)
Doctorate in applied mathematics and computer science (1979)
28-year career as an engineer and consultant in various sectors (1980-2007) (civil engineering, medical imaging, defense, telecoms, air traffic management systems)
Reconversion to teaching and linguistics (2007-2009)
CAPES and bi-admissibility for the agrégation d'anglais, linguistics option.
8-year career as a specialist English teacher at the University Paris-Diderot University (2009-2017)
Retired since 2017.
Personal interests in creative activities:
writing (literaryliterary fiction, translation essays)
poetry translation (from English, Spanish, Italian and Russian into French)
chess
- 17h30 Closing