A research position for one Bolsa de Investigação para Mestre (1 BI) is opened at Instituto de Telecomunicações, in the scope of the Project MAIA ? Multilingual AI Agents for Customer Service ? Projeto Parceria Internacional CMU ref. 045909, funded by the applicable financial framework. The work for this position is in the area of Machine Learning and Natural Language Processing. The candidate should have a Master's degree and enrol in a PhD program. In the scope of the project may visit Carnegie Mellon University (a partner in the project).
Scientific Area:
Information and Data Sciences
Group:
Pattern and Image Analysis ? Lx
Admission Requirements:
MSc in Electrical and Computer Engineering, Informatics, Data Science, Mechanical/Aerospace Engineering, Applied Mathematics, or equivalent scientific areas - Excellent academic and practical background in programming, machine learning, and natural language processing - The candidate should have good knowledge of written and spoken English, be self-motivated and show initiative in solving problems - Strong preference is given to candidates with immediate availability to initiate the work and sound knowledge of Python and Pytorch.
Work Objectives:
The goal of MAIA is to develop multilingual AI agents for customer support. The work done in this project will be developed at IT in collaboration with Unbabel, CMU, and INESC-ID. There are three tasks (related to Activity 3 in the project): - Develop new latent structured models for discrete data. These models should cope with discrete latent variables or generate discrete data. This will allow an efficient and compact representation of the context of the conversation, leading to compact and data-efficient models, with fewer parameters than current state-of-the-art neural network models for language generation. These models will be applied to machine translation and response generation. - Investigate methods based on sparse attention mechanisms to achieve explainability and return confidence scores about models' predictions. This will help human agents that work collaboratively with language generation systems, and, in machine translation, it will enable the use of active learning strategies to decide which data to request human translations for supervision. These methods will be used to incorporate conversational feedback and for conversational quality estimation. - Investigate methods for reducing the need for labeled data. These methods will leverage unlabeled data, such as monolingual text, which will used to pre-train parts of the network and help the target task via transfer learning. In addition, we will explore connections to the previous task by using latent variables to learn representations form the unlabeled data
Applicable legislation:
A fellowship will be celebrated according to the regulations defined by FCT, ?Regulation of Instituto de Telecomunicações? fellowships? and ? Estatuto do Bolseiro de Investigação?, according to Law no. 40/2004, dated August 18 (Status of Scientific Research Fellow), in its actual redaction, and also by Fundação para Ciência e Tecnologia, I.P. Fellowship Regulation of Fundação para Ciência e Tecnologia, I.P.
Place of work:
The work will be developed at the premises of Instituto de Telecomunicações at Lisbon, under the supervision of Professor André Martins.
Duration:
12 months, eventually renewed in accordance FCT scholarship regulations.
Monthly salary:
1074.64 ? (net), according to the table of grant amounts awarded directly by FCT for positions held in Portugal (www.fct.pt/apoios/bolsas/valores). The payment will be made by bank transfer.
Selection criteria:
----