Hi! I am Pietro Lesci, a senior associate in data science at Bain & Company in Milan, Italy. While working towards my Msc in Economic and Social Sciences at Bocconi University, I have worked as a trainee in data science at the European Central Bank in Frankfurt am Main, Germany. After graduation, I started collaborating as a research assistant in Natural Language Processing at the Bocconi Institute for Data Science and Analytics with Prof Dirk Hovy in Milan, Italy. As of March 2020, I am working as a data scientist in the Advanced Analytics Group at Bain & Company.
First of six children, I am a passionate musician and a statistics enthusiast.
Today’s world is full of acronyms! — Worry not, I’ll unpack them shortly.
I am interested in meta-learning (ML) and active learning (AL) in the context of natural language processing (NLP) — three acronyms explained, we’re still missing one ML, though. NLP is a huge field of research and, as such, there exist many “angles” from which it is possible to study it (e.g., linguistics, machine learning, neuro-science, etc) — small hint on the second ML there. I am passionate about the probabilistic machine learning (ML) approach to NLP — here you go with the second ML to “complete the square”.
I developed this passion while working towards my MSc thesis with Prof Sonia Petrone. The MSc at Bocconi University was centered on Bayesian methods for computational social sciences and, most importantly, Bocconi University is a renowned center for Bayesian nonparametrics. Influenced by this exciting context, in my thesis I proposed a Bayesian interpretation of a class of meta-learning algorithms called Neural Processes. This project allowed me to familiarize myself with the literature on deep latent variable models, Bayesian deep learning, and their implementations via probabilistic programming languages.
At the Bocconi Institute for Data Science and Analytics, I was a member of the team that is currently working on a Twitter funded project to understand, explore, and measure the health of conversations on the platform. In particular, while on the project, my focus was the identification of echo chambers and online abuse focusing on methods for the inclusion of covariate information in word-embeddings. I have worked with NLP algorithms and performed data analytics on textual data.
Despite working at the pace of a top-tier consulting firm at Bain & Company (that decisively stress-tested my passion for research), I try to dedicate my free time to pursuing my research interests. In fact, I am continuining my collaboration with Prof Dirk Hovy’s lab at Bocconi University focusing on meta-learning and active learning in NLP. As part of this collaboration, I am having the amazing opportunity to work with professors and PhD students from all over the world.
Having a statistical (vs a computer science) background taught me the hard lesson that, especially nowadays, stats theory and R scripting is not enough to keep up with the ML^2 (see section above) literature. That’s why during my MSc studies and while at the European Central Bank in 2018, I’ve started focusing on acquiring engineering skills (e.g., software development principles, collaboration via Git/Github, Docker, Kubernetes, etc) — effort that I am still pursuing on a daily basis as part of my current job as a data scientist. Ultimately, this effort paid off in many ways: getting me a job in data science/engineering, allowing me to quickly implement research ideas, and being able to collaborate on open-source projects.
At Bain & Company, I am a core contributor on the development of an internal Python library for text analytics. Also, I’ve lead a large-scale experimentation project with the aim to solve a challenging textual classification task for a client. While on this project, I’ve had the chance to implement and test many state-of-the-art NLP models.
As part of my work at Bocconi University, I deployed two machine learning web-apps — Wordify and MACE — that allow researchers in other fields (e.g., social sciences, marketing, economics, etc) to easily access NLP tools.
MSc in Economic and Social Sciences, 2019
BSc in Economics and Management, 2016
Università Cattolica del Sacro Cuore
When evaluating redundant annotations (like those from Amazon’s MechanicalTurk), we usually want to (i) aggregate annotations to recover the most likely answer, (ii) find out which annotators are trustworthy, (iii) evaluate item and task difficulty. MACE solves all of these problems, by learning competence estimates for each annotators and computing the most likely answer based on those competences.
Wordify is an online textual processing tool that allows to find out which terms are most indicative for each dependent variable values.