Marco Gaido received the Best Student Paper Award at IWSLT 2022

Marco Gaido, a PhD Student at the Machine Translation Research Unit, received the Best Student Paper Award at the International Conference on Spoken Language Translation (IWSLT 2022)

In particular, the research paper focuses on the problem of recognising and properly translating specific entities (person names) present in speech data.

His solution includes the creation of models capable to handle a wide range of phonetic features, thus being robust to the mispronunciation of names of people from different nationalities.

From the paper conclusions: “Humans and machines have different strengths and weaknesses. Nonetheless, we have shown that when it comes to person names in speech, they both struggle in handling names in languages they do not know and names that they are not used to hear. This finding seems to insinuate that humans cannot expect help from machines in this regard, but we demonstrated that there is hope, […] Indeed, since machines are faster learners than humans, we can train them on more data and more languages. Moreover, we can design dedicated architectural solutions aimed to add an inductive bias and to improve the ability to handle specific elements.”

“We are particularly proud of this result” – says Matteo Negri, co-author of the paper together with Marco Turchi -. “The underlying research raises the bar in speech translation, opening up to new avenues where translating means much more than correctly rendering speech content in a different language. Indeed, our ongoing efforts aim at the more ambitious objective of enhancing the output of our systems with semantic information – in this case, the annotation of named entities present in a speech – suitable for a variety of downstream tasks. Towards this objective, which lies at the crossroads between speech translation and natural language processing, the award represents a first important milestone for both Marco and for our group.”