NLP at Cleverly and Multilingual Email Zoning

1 June 2021

Mariana Almeida Cleverly

Cleverly is an end-to-end AI layer for customer service platforms that provides intelligent automation and efficiency and unlike others is easy to use. We reduce the effort agents spend on repetitive tasks and searching for the right information, giving them more time to focus on the complex queries. In this talk, we will show how Cleverly tackles Customer Service productivity issues. Then, we will focus on the research work Cleverly did on Multilingual Email Zoning, and we recently presented at EACL 2021. As a basic component of the email NLP pipeline, Cleverly has an email zoning module that identifies the semantic zones in an email, such as “greetings”, “body”, “signatures”, “attachments”, etc. However, and despite most email applications being multilingual, pre-existing literature on email zoning is limited to English corpora and models. Considering this, we created Cleverly Zoning Corpus (the first multilingual corpus available for email zoning) and developed OKAPI, a multilingual email zoning model. The results show that OKAPI generalizes well for unseen languages. Our zoning model also achieves competitive performances in English and reaches a new state-of-the-art on domain adaptation tasks in English.



NLP at Cleverly and Multilingual Email Zoning banner Mariana Almeida leads Research and Data Science at Cleverly, since December 2019. There she helps to drive NLP (Natural Language Processing) and ML (Machine Learning) to make customer support teams more efficient. Before that, she worked as a researcher for 12 years both in academia and in the industry. She spent 4 years at Priberam Labs exploring ML and NLP methods and 2 years at Feedzai working on ML and deep learning methods for fraud prevention. In academia, she worked on image processing: she was a Postdoc fellow at IT of Lisbon, working on image reconstruction and optimization under the supervision of Mário A. T. Figueiredo; and she did a PhD in blind image separation and blind deconvolution, under the supervision of Luís B. Almeida. Mariana received the Vidigal award 2004/2005 and an honourable mention at the 2010 Portuguese IBM Scientific Prize. She is also co-organizer of LxMLS since 2016.