Research Archive

Natural Language Processing

These pages document some of the natural language processing tools that we are using to develop a digital representation of the Sicilian language at Napizia. Specifically, the tools that use machine learning.

Recent research in the field has focused on making neural machine translation work for low-resource languages (i.e. languages for which little parallel text is available). Sicilian could make a good case study for those efforts because Arba Sicula has published an enormous quantity of Sicilian-English translations in its 40-year history. But prior to my work on this project, no one had ever assembled it into a dataset of parallel text.

So we have a large quantity of material to assemble into a dataset. And as we do so, we can test some of the methods being used to bring neural machine translation to low-resource languages.


Sicilian Language

Sicilian is a beautiful language. To bring its beauty to the world, we're developing a machine translator for the language and we're annotating the Dieli Dictionary with poems, proverbs and prose.

Arba Sicula and Arthur Dieli have kindly supplied the translations necessary to assemble the dataset of parallel text necessary to develop a machine translator. And they have provided a lot of support and encouragement. So we hope to bring a good quality translator online soon.

We are also encouraged by the work of Sennrich and Zhang (2019) who obtained large improvements in translation quality with the method of subword splitting. That method yielded similar improvements with the Sicilian-English dataset that we are assembling, so we hope to bring a good quality translator online soon.

In the meantime, you can see our progress by visiting the neural machine translator that we posted online. And you can discover the beauty of the Sicilian language by reading the poetry, proverbs and prose in Cchiù dâ Palora.


Labor Law

After the 2008 Financial Crisis triggered a global recession, policymakers in many countries have attempted to increase employment by making their labor market "more flexible." The specific proposals differ by country, but they share the common theme of reducing worker protections.

In France, reforms that would allow companies to lay off workers more easily, loosen restrictions on working hours, reduce overtime payments and reduce severance payments sparked the Nuit debout protests in the spring of 2016.

Italy's experience with labor market liberalization (following passage of the "Biagi Law" in 2003) suggests that such reforms will not increase employment rates, because macroeconomic growth affects employment rates far more than labor market regulation does.

In the United States, states with higher minimum wage rates tend to have higher employment rates. They also tend to have higher average annual pay.

More broadly, OECD countries with stronger protection against dismissal tend to have higher employment rates. And OECD countries with tighter regulation on temporary forms of employment tend to have higher male employment rates, but lower female employment rates.


AnX11 Phone Project

This documentation project explains how to install and configure Debian on a rooted Android phone. Installing desktop software onto a phone provides a nice "on the go" solution and it's a relaxing way to work.

As an economist with Debian installed on my Android phone, I can run R scripts and Perl scripts directly on the phone itself. And, if I'm teaching class, I can project X11 applications like wxMaxima and Gretl from my phone onto the screen behind me. Students may follow my lead on their laptop computers. Or -- if students also have Debian installed on their phones -- they can follow my lead on their phones.


Mortgage Default

In 2010, the NYS Banking Dept. (NYSBD) began collecting data on every home mortgage default and foreclosure in the New York State for the purpose of transmitting it to non-profit mortgage counselors. We paired the NYSBD's data with data on originations from the Home Mortgage Disclosure Act (HMDA) enables us to compare borrowers who defaulted to those who did not.

Like many previous studies, we found strong racial and ethnic disparities in lending practices and our analysis shows that the same disparities reappear in the pre-foreclosure filing data. This finding suggests that lending disparities contributed to the higher default rates that we observe among black and Latino borrowers.

Employment growth appears to have the strongest effect on the home mortgage default rate. In the absence of employment growth, even large principal balance reductions would only have a minimal effect on the rate of mortgage default.



While working at the NYSBD, Sean MacDonald and I developed coincident indices and used those indices to forecast the state of the NYS labor market and the financial health of NYS banks.


Income Inequality

In the economic development literature, much work is devoted to the question of whether income inequality will widen or narrow as an economy grows over time. Most models assume inequality, but in this paper I developed a model that is both grounded in growth theory and capable of generating all degrees of income inequality (from complete equality to complete inequality).


Health Outcomes

This is a feasibility report on case management partnerships between New York City's HIV-AIDS Services Administration and the Health and Hospitals Corporation. The two organizations collaborated to help mutual clients persist in medical and behavioral health care and to meet the long-term housing needs of mutual clients.

By creating a formal working relationship between HASA and HHC-COBRA programs, the pilot eliminated duplication of effort, kept clients connected to medical and behavioral health care and helped clients who needed to relocate avoid emergency housing.

Copyright © 2014-2019 Eryk Wdowiak