Artificial Intelligence and Learning Systems Laboratory

Panagiotis Kouris, Γιώργος Αλεξανδρίδης and Andreas-Giorgos Stafylopatis work on text summarization based on semantic graphs has just been published open access by the Journal of Big Data by SpringerOpen!

Text summarization based on semantic graphs: an abstract meaning representation graph-to-text deep learning approach - Journal of Big Data Nowadays, due to the constantly growing amount of textual information, automatic text summarization constitutes an important research area in natural language processing. In this work, we present a novel framework that combines semantic graph representations along with deep learning predictions to g...

Our joint work with Deeplab on Self-Supervised Learning for Visual Relationship Detection through Masked Bounding Box Reconstruction has just been presented at the Winter Conference on Applications of Computer Vision (WACV) 2024 as a poster!

WACV 2024 Open Access Repository

📝Want to use a counterfactual editor but don't know if its edits are truly minimal? Measuring inconsistency may help!
Our paper “Counterfactuals of Counterfactuals: a back-translation-inspired approach to analyse counterfactual editors”, delves deep into this.

🔍 We introduce a novel metric which uses iterative feedback steps to evaluate the inconsistency of editors!

🔍 The behavior and outputs of counterfactual editors varies a lot, but there is no universal ground truth for counterfactual edits. As such it is hard to tell what would be the optimal counterfactual.

🔍 We propose a using previous editor outputs as ground truth proxies.

🔍 We compare three editors on two datasets covering different types of use cases (adversarial and counterfactual, black and white box, with or without LLMs), gaining useful insights on the editors’ behavior.

🔍 Some insights : MiCE achieves lowest minimality but tends to leave remnant text spans, as indicated in the example below. This observation is also reflected in the inc@n: there is a higher value of inc@n when n is even indicating that it is easier to return to the original class.

📝 You can checkout the paper here: https://arxiv.org/abs/2305.17055 or talk to us in person at (come to our spotlight talk too).

w/ Eddie Dervakos, Orfeas Menis, Chryssa Zerva and George Stamou

Ο Τομέας Τεχνολογίας Πληροφορικής και Υπολογιστών της Σχολής Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών του Εθνικού Μετσόβιου Πολυτεχνείου, και το Διατμηματικό Πρόγραμμα Μεταπτυχιακών Σπουδών «Επιστήμη Δεδομένων και Μηχανική Μάθηση» σας προσκαλούν σε επιστημονική ημερίδα με θέμα τις σύγχρονες ερευνητικές προκλήσεις στην Επιστήμη Υπολογιστών. Η ημερίδα θα γίνει την Δευτέρα 9 Ιανουαρίου 2023, στην Αίθουσα Τελετών, στο ισόγειο του Κτηρίου Διοίκησης, στην Πολυτεχνειούπολη Ζωγράφου, σύμφωνα με το παρακάτω πρόγραμμα.

Computer Science Talks, 9 January 2023, 16:00-20:00, Conference Hall, NTUA Administration Building Vassilis Zikas, Vasiliki (Vasia) Kalavri, Constantine Caramanis, Manolis Zampetakis