T2KG: Transforming Multimodal Document to Knowledge Graph

Proceedings of Recent Advances in Natural Language Processing, pages 385–391 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_043 385 T2KG: Transforming Multimodal Document to Knowledge Graph Santiago Galiano and Rafael Mu˜noz, and Yoan Guti´errez and Andr´es Montoyo and Jose I. Abreu Research Group of Language Processing and Information System. University of Alicante sgs97@alu.ua.es, {rafael,ygutierrez,montoyo,jabreu}@dlsi.ua.es L. Alfonso Ure˜na University of Ja´en laurena@ujaen.es Abstract The large amount of textual information, in dig- ital format available today, makes the knowl- edge extraction task unfeasible by manual means. It is therefore necessary to develop automatic tools that allow us to integrate this knowledge into a structure that is easy to use by both machines and humans. This paper presents T2KG, a framework that can incor- porate the relevant information from several structured or unstructured documents into a se- mantic network. Structured documents are pro- cessed based on their annotation scheme. For unstructured documents, T2KG uses a set of Natural Language Processing sensors that iden- tify relevant information to enrich the semantic network created by linking all the knowledge from different documents. 1 Introduction Nowadays, the amount of information available on the web is available in multiple formats. Leverag- ing this data requires the design of software sys- tems that can exploit the information, obtain rele- vant data, structure it in a specific format, and gen- erate reports that help to evaluate this information. Software systems focused on performing all these actions are currently oriented to apply different nat- ural language processing techniques. Many early developments were domain-dependent, so domain- specific resources, although costly in terms of time and expertise, were relatively easy to obtain. But recently, general-purpose, domain-independent sys- tems are being developed. However, it would be difficult to imagine a system like ChatGPT incor- porating a multi-domain ontology in real time. The trend in Natural Language Processing (NLP) today is text-to-text development that does not use man- ually curated semantic resources such as seman- tic networks. In other words, text-to-text oriented systems use self-generated resources without the need for external semantic resources. However, systems must take into account that the software created must be maintainable and extensible, using processes and methodologies that make all these aspects possible. This work aims to present a framework capa- ble of extracting knowledge from heterogeneous sources, structured such as comma-separated vol- umes or relational databases, or unstructured such as plain text from Wikipedia. Knowledge from dif- ferent sources is integrated into an ontology. It also allows the user to query the knowledge in natural language while results are analyzed automatically to generate custom graphics or visualizations to ease its interpretation. 2 State of the Art Knowledge representation is the process of mod- eling information in a way that enables effective reasoning, communication, and decision-making by computers. Given the increasing amount of digital data available, it has become more impor- tant than ever to conceive ways to represent it in a meaningful way to add knowledge to NLP sys- tems. This knowledge has been used to improve the accuracy of tasks …

Content truncated to 500 words