Abstract
This thesis belongs to the research field of smart data, where we search for specific information within textual documents. It proposes new methods of representation and extraction of experimental data from scientific articles. These methods were evaluated on a corpus of articles in the food packaging domain.The experimental data can be represented as n-Ary relations composed of symbolic and quantitative arguments. The latter are composed of a numerical value and a unit of measurement. The objective of this thesis is to populate a knowledge base with instances of N-Ary relations extracted from scientific textual documents. The proposed approach is based on an Ontological and Terminological Resource (OTR) and is divided into two Phases: (1) the recognition and extraction of argument instances of interest and (2) the linking of these instances in n-Ary relations. Phase (1) proposes an original representation of the extracted argument instances, called SciPuRe (Scientific Publication Representation). It integrates ontological, lexical and structural descriptors that describe the context of the argument instances and allows to sort them by their relevance. Phase (2) relies on the information present in the tables of the documents, extracted automatically, to guide the extraction of partial n-Arye relations, the tables containing an important part of the experimental data in the scientific articles. These partial relations are then completed with the argument instances recognized in Phase (1). Three approaches are proposed and evaluated in order to identify the argument instances that should complete the relations: the use of document structure, the analysis of cooccurrences between the argument instances in the texts, and finally the use of word-embedding models allowing to measure the similarities between the candidate argument instances and the arguments already filled in the partial relations.Our results show the importance of sorting the relevant instances after argument recognition in Phase (1) using SciPuRe features. Our experiments show that the two most important criteria for determining the relevance of a symbolic argument instance are the specificity of the concept associated with the argument in the OTR and its frequency in the document. For quantitative arguments, it is the apparition of the argument instance in sections of the documents that determines its relevance. Our experiments on Phase (2) confirm the usefulness of the relevance scores computed in Phase (1) to discriminate the instances. The analysis of the results with different filtering of the candidate argument instances according to their relevance shows a clear positive effect when filtering 20%$ of the instances with the lowest relevance. We also experimented with the possibility of selecting multiple candidates for each missing argument instance in a partial relation, in an approach to assist domain experts who can then determine the valid instance. When selecting a single candidate, the approach based on co-occurrence analyses gives the best results in detecting the valid candidate argument instance. With a larger selection of three or five candidates, semantic similarity analysis enabled by BERT word embeddings model provides good results for detecting associations between the argument instances present in partial relations and the candidate argument instances for relation completion. Finally, when selecting ten candidates, the experiments show that the approach based on document structure is effective to complete the n-Ary relations.