#WiNLP author spotlight no. 11 Marcely Zanon Boito - "Unsupervised Word Discovery Using Attentional Encoder-Decoder Models"

===========

"My name is Marcely Zanon Boito and I was born in the city of Porto Alegre, Brazil. I’m an exchange student from the Federal University of Rio Grande do Sul (UFRGS, Brazil) and Institut National Polytechnique de Grenoble (Grenoble-INP, France), and I’m currently working in the GETALP (Groupe d’Étude en Traduction Automatique/Traitement Automatisé des Langues et de la Parole) research group from Grenoble, with the collaboration of the NNLP (Neurocognition and Natural Language Processing) research group, from UFRGS.

I’ve always been deeply interested in the integration of human aspects to science, and that’s why I consider the NLP research field so appealing. In particular, I find the different challenges presented in information extraction for low-resource/endangered language scenarios really interesting, since we do not have a lot of available information to help us.
I had my first experience in NLP in my first year as an undergraduate student in Computer Science, working as a research assistant in the NNLP research group. There I participated in a two-years project focused in lexical simplification.

Right now I’m interested in understanding the possible contributions that neural-based approaches could have in two fields: language acquisition and documentation. This investigation includes the tasks of word and lexicon discovery in unsupervised and semi-supervised scenarios, and the abstract presents our preliminary results for the task of unsupervised word discovery for language documentation.

I chose to submit to the WiNLP workshop because I believe it will be an empowering experience to see women and other minorities having their research highlighted and having their voices heard in a safe environment. I’m eager to get to know the research topics and contributions of all these amazing people in the NLP field."

===========

Marcely will be presenting a poster in the lunch poster session, 13 - 14:30 on Sunday, 30th July 2017.
Poster abstract:

In this project we explore soft-alignment probability matrices generated by an attention-based sequence-to-sequence neural machine translation system, in order to evaluate if these soft-alignments allow us to discover latent lexicon representation. Our goal is to understand if it is possible to use these soft-alignments for unsupervised segmentation and lexicon discovery in low-resource scenarios, limited by the amount of data available.