Combiner espaces sémantiques, structure et contraintes.

This paper presents the methods that we developed for the tasks 1 and 4 of the DEFT'14 Text Mining contest. In the task 1 the goal was to automatically categorise the literary genre of short texts, while in the task 4 the goal was to assign the session where a scientific paper is presented in a conference by analysing its content. These methods we developed rely on a common representation of the input texts in semantic spaces constructed using Random Indexing. In these high dimension spaces, each text and each term is represented a vector. For this edition of the DEFT, we tried to address the proposed tasks by designing methods that combine classical machine learning algorithms for clustering and categorisation with (i) rule based methods to represent for instance the patterns of poetic texts in the task 1 (ii) constraints solving methods to take into account the informations we had about the organisation of the sessions in the task 4. The results obtained NDCG=0.4278 (rank 2) in the task 1 and FScore=1 (rank 1) in the task 4 show the great performance of these hybrid methods.

EL GHALI Adil;  EL GHALI Kaoutar; 

2014-09-10

ATALA

JRC90537

http://deft.limsi.fr/actes/2014/pdf/deft2014_02_lutin.pdf,    https://publications.jrc.ec.europa.eu/repository/handle/JRC90537,