Lundi 11 Mars
Heure: 
14:00  15:00 
Lieu: 
Salle B107, bâtiment B, Université de Villetaneuse 
Résumé: 
Structure Prediction Energy Network (SPEN) using Dual Decomposition on Dependency Parsing 
Description: 
Xudong ZHANG Dependency Parsing is one of the basic tasks in the field of Natural Language Processing (NLP). The goal is to find whether there exist a strong relationship between different words in a sentence. It can be used as the basic step of many NLP systems like question answering system. Solving a dependency parsing problem can be realized by a energy based network with the output of the neural network as a scalar (energy). The goal is to find the most compatible structure (a graph) with the input sentence and the most compatible structure is supposed to give the lowest energy for the neural network. As the structure of the sentence should be a tree (one root, every word has and only has one pa rent, no circle), to simplify the problem, people always construct a linear function corresponding to the structure that we want to find, i.e. we suppose different arcs are independent. However, this method may limit the capacity of the system to describe more complex relations. In this project, inspired by the idea of Structure Prediction Energy Network (SPEN), we construct a new neural network which is composed of two parts, i.e. local energy part and global energy part. We showed that it is possible to solve the problem with dual decomposition when we have a convex (nonlinear) function for the global energy part together with the linear local energy part. As one part of my Phd thesis, this work is still ongoing. 

