During the last decade, the need for reliable and massive Knowledge Graphs (KG) increased. KGs can be created in several ways: manually with forms or automatically with Information Extraction (IE), a natural language processing task for extracting knowledge from text. Relation Extraction is the part of IE that focuses on identifying relations between named entities in texts, which amounts to find new edges in a KG. Most recent approaches rely on deep learning, achieving state-ofthe-art performances. However, those performances are still too low to fully automatize the construction of reliable KGs, and human interaction remains necessary. This is made difficult by the statistical nature of deep learning methods that makes their predictions hardly interpretable. In this paper, we present a new symbolic and interpretable approach for Relation Extraction in texts. It is based on a modeling of the lexical and syntactic structure of text as a knowledge graph, and it exploits Concepts of Neighbours, a method based on Graph-FCA for computing similarities in knowledge graphs. An evaluation has been performed on a subset of TACRED (a relation extraction benchmark), showing promising results.

Knowledge Graphs (KG) offer easy-to-process information. An important issue to build a KG from texts is the Relation Extraction (RE) task that identifies and labels relationships between entity mentions. In this paper, to address the RE problem, we propose to combine a deep learning approach for relation detection, and a symbolic method for relation classification. It allows to have at the same time the performance of deep learning methods and the interpretability of symbolic methods. This method has been evaluated and compared with state-ofthe-art methods on TACRED, a relation extraction benchmark, and has shown interesting quantitative and qualitative results.

Language registers are a strongly perceptible characteristic of texts and speeches. However, they are still poorly studied in natural language processing. In this paper, we present a semi-supervised approach which jointly builds a corpus of texts labeled in registers and an associated classifier. This approach relies on a small initial seed of expert data. After massively retrieving web pages, it iteratively alternates the training of an intermediate classifier and the annotation of new texts to augment the labeled corpus. The approach is applied to the casual, neutral, and formal registers, leading to a 750M word corpus and a final neural classifier with an acceptable performance.