LCT Annual Meeting 2024

The Annual Meeting of the Erasmus Mundus European Masters Program in Language and Communication Technology (LCT) brings together current students, graduates, and faculty members from partner universities for two days of learning, networking, and celebration. The event features talks from invited speakers (including the alumni of the program), a dynamic poster session showcasing student research, and a graduation ceremony honoring our latest graduates. Attendees will have the opportunity to network with peers and professionals during breaks and a joint dinner. This meeting fosters collaboration, highlights academic and professional achievements, and brings the LCT community together.

Agenda Day 1 - Monday, June 17, 2024

08:45 - 09:30 Registration
Location: ground floor, Room 10

09:30 - 10:00 Welcome speech by:
Uri Hasson from the CiMeC Board
Josef van Genabith, LCT project coordinator
Evgeniya Ustinova, LCT project management 

10:00 - 10:45 Invited speaker
Chair: Jakub Szymanik, CIMeC

It takes two to TANGO: towards human-machine learning and decision making
Andrea Passerini DISI, University of Trento

In this talk I will argue that only a deep mutual understanding between humans and machines can allow them to jointly achieve results that are far better than those that each would achieve independently. I will provide a comprehensive overview of the research activities undertaken by the TANGO project, highlight some of our promising results, and discuss relevant research directions aimed at developing truly human-centric AI systems.

10:45 - 11:15 Coffee break
Location: ground floor, Rooms 8 and 9

11:15 - 13:15 Parallel meetings
Location students: Aula Magna (Palazzo Piomarta) - only offline
Location consortium & members of the EQAB: Aula Magna (Palazzo Fedrigotti)

13:15 - 14:15 Lunch
Location: ground floor, rooms 8 and 9

14:15 - 14:55 Poster session (Group 1) Location: ground floor 
15:05 - 15:45 Poster session (Group 2) Location: ground floor 

15:45 - 16:30 Coffee break

16:30 Graduation Ceremony

Joint picture

19:30  Joint dinner at Moja Ristorante Caffetteria

Agenda Day 2 - Tuesday, June 18, 2024

09:00 - 09:30 Second day registration
09:30 - 10:15 Invited speaker

Bias in Speech Processing
Ajinkya Kulkarni, Idiap, LCT intake 2016

Advances in speech technologies have significantly improved tasks like Automatic Speech Recognition (ASR) through models such as Whisper and Multilingual Massive Speech (MMS). However, there remains a critical need to address and understand the biases these systems can perpetuate. This presentation explores the complexities of studying biases in speech processing applications, focusing on Whisper and MMS systems. It examines the intersection of bias with factors such as voice timbre, skin tone, and age in a multilingual ASR setting. Additionally, the presentation delves into the environmental impact of these ASR systems, considering carbon emissions and energy consumption. By highlighting these issues, the presentation underscores the necessity for ongoing research into bias and sustainability in speech-processing technologies.

10:15 - 11:15 Plenary meeting: consortium & students
Chair: Markéta Lopatkova

11:15 - 11:45 Coffee break
11:45 - 12:30 Alumni session. Graduates of the intake 2021 will share their experience after finishing the program.

12:30 - 13:30 Lunch

13:30 - 14:30 Industry session
Chair: Marc Tanti

13:30 Introduction
13:40 Georg Rehm, DFKI (online)
13:50 Arantza del Pozo, Vicomtech (online)
14:00 Carlo Strapparava, FBK (online)
14:10 Discussion of issues related to the internship

14:30 - 15:00    Invited speaker (hybrid)

When Your Cousin Has the Right Connections: Unsupervised Bilingual Lexicon Induction for Related Data-Imbalanced Languages (best student paper award at LREC-Coling 2024)
Niyati Bafna, Johns Hopkins University, LCT intake 2020

Most existing approaches for unsupervised bilingual lexicon induction (BLI) depend on good quality static or contextual embeddings requiring large monolingual corpora for both languages. However, unsupervised BLI is most likely to be useful for low-resource languages (LRLs), where large datasets are not available. Often we are interested in building bilingual resources for LRLs against related high-resource languages (HRLs), resulting in severely imbalanced data settings for BLI. We first show that state-of-the-art BLI methods in the literature exhibit near-zero performance for severely data-imbalanced language pairs, indicating that these settings require more robust techniques. We then present a new method for unsupervised BLI between a related LRL and HRL that only requires inference on a masked language model of the HRL, and demonstrate its effectiveness on truly low-resource languages Bhojpuri and Magahi (with <5M monolingual tokens each), against Hindi. We further present experiments on (mid-resource) Marathi and Nepali to compare approach performances by resource range, and release our resulting lexicons for five low-resource Indic languages: Bhojpuri, Magahi, Awadhi, Braj, and Maithili, against Hindi.

15:00 Closing