Graph-based knowledge representation and extraction from unstructured textual data using machine learning algorithms
Loading...
Date
2023-07-21
Authors
Κανακάρης, Νικόλαος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
In this dissertation, we build upon existing graph-based text representation models and introduce a novel one to represent multiple textual documents as a single graph, namely, the graph-of-docs text representation. As opposed to existing graph-based text representations, graph-of-docs enables the investigation of the importance of a term into a whole corpus of documents, masks the overall complexity by reducing each graph of words to a ‘document’ node, and supports the inclusion of relationship edges between documents. Hence, it enables the calculation of important metrics with respect to multiple documents and allows heterogeneous data to co-exist in a single graph. Along with the introduction of the graph-of-docs model, this dissertation proposes and empirically evaluates the combination of word embeddings, graph-based text representations and graph neural networks to advance classical machine learning tasks, including text classification, regression, feature engineering and feature selection. We conduct several experiments on diverse datasets from different domains, settings and applications, aiming to evaluate the proposed approach. Among others, these datasets are related to (i) personnel selection, (ii) the identification of software bugs, (iii) the prediction of future research collaborations, (iv) the price prediction for Airbnb listings. The evaluation results demonstrate a significant improvement in terms of accuracy, precision and recall of the proposed models compared to various classical and graph-based counterparts.
Description
Keywords
Graph mining, Machine learning on graphs, Graph-based text representations, Graph neural networks, Natural language processing, Link prediction, Node regression, Graph and node classification