Athens University Economics and Business (AUEB)
Modeling and extracting research processes
Abstract
dc:descriptionWe define a conceptualization of the research process, the Scholarly Ontology (SO),based on which we show how different aspects of scholarly research behaviour canbe documented automatically as well as interactively and relevant resources, possiblyfrom different fields, can be interconnected. We demonstrate the capacity of SO tosupport various research inquiries by formulating appropriate complex queries. Weexplore the role of type taxonomies, such as the Activity Type class, and explain howtheir instances can function as pivotal elements in modeling intentionality andfunctionality properties of various SO classes. We formulate semantic constraintsthat provide a framework for reasoning over SO. Furthermore, we investigate the useof semantic similarity in process classification via the Activity Type taxonomy. Wereview several graph-based similarity measures and we propose a new, highperformance one based on a novel definition of semantic specificity that leveragesstructural properties of graphs such as WordNet. We then turn to the creation of aknowledge base by information extraction from text, driven by SO. We do this usingtwo approaches: (1) devising specialized NLP rules that fully exploit the ontologyand structural and syntactic properties of scientific publications; and (2) usingstatistical machine learning methods. Along the first line we developed ResearchSpotlight, a system that leverages information from DBpedia, retrieves articles fromvarious sources (e-repositories, Web pages), extracts and interrelates various kindsof named and non-named entities by exploiting article metadata, the structure of textas well as syntactic, lexical and semantic constraints, and populates a knowledgebase in the form of RDF triples. Specialized rules were defined, implemented andembedded in the outputs provided by SPACY, an industrial-strength framework forNLP tasks such as tokenization, segmentation, POS tagging and dependency parsing.Evaluation of the system is performed through various experiments (token-based, entity-based) that measure the overall performance as micro / macro-averagedPrecision, Recall and F1 scores of the classes and relations that are automaticallypopulated. We also perform error analysis, to investigate the source of potentialerrors and acquire insights for future improvements. Finally, we explore the use ofstatistical machine learning methods in extracting research processes from text, withfocus on populating three core classes of SO: the unary predicate Method comprisingnamed entities that denote procedures in scholarly context, the unary predicateActivity -denoting research processes- and the binary predicate follows(act1, act2) -denoting sequential relations between activities. We present, in total, 13 slidingwindow machine learning methods for token, sentence and chunk-basedclassification employing Logistic Regression, linear Support Vector Machines andRandom Forests, which incorporate word, POS and dependency embeddings, as wellas combinations of handcrafted features that leverage the syntactic, lexical andstructural attributes of the texts. Evaluation in all cases, shows promising results,way above the baselines, encouraging the use of similar methodologies for theextraction of other relevant concepts and relations in the wider domain of scholarlypublications, as well as for experimenting with other more complex classifiers (e.g.convolutional or recurrent neural networks) when adequate datasets becomeavailable.
Degree
thesis:*- Grantor dc:publisher
- Athens University Economics and Business (AUEB)
- Year dc:date
- 2018
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Pertsas, Vayianos
- Περτσάς, Βαγιανός
Subjects
dc:subject × 16- Οντολογίες
- Μηχανική μάθηση
- Εξαγωγή Πληροφορίας από Κείμενο
- Επεξεργασία φυσικής γλώσσας
- Εννοιολογική μοντελοποίηση
- Domain ontology
- Machine learning
- Natural language processing
- Information extraction
- text minning
- Conceptual modeling
- process modeling
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Natural Sciences
- Computer and Information Sciences
Rights
dc:rights- Statement dc:rights
-
- BY_NC_ND
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/43604