Résumé
During this last decade, word embeddings models learning continuous vector representations of words have been established and integrated in several applications of Natural Language Processing (NLP). These models have been subsequently extended to learn representations of other textual objects such as word senses/definitions, fragments of textual documents or even whole texts. In this paper, we focus on the creation of continuous vector representations for relations. We propose a model where vectors embeddings of relations are deduced from (multi-)words embeddings. The training of these representations is carried out from a business corpus referring to several specialized domains such as health, justice, urbanism, or elections. The quality of these representations is evaluated on the task of identifying/classifying lexical-semantic relations from texts with binary classifiers (Machine Learning techniques). This task consists in classifying from a relation embedding in input, the relation type of this representation. The obtained results are good and surpass the performances for a recent one state-of-the-art system dedicated to the creation of relations representations.