The project T2Know presents the use of natural language processing technologies for the creation of a semantic platform of scientific documents via knowledge graphs. This knowledge graph will link relevant parts of each document with those of other documents in such a way that trend analysis and recommendations can be achieved. The goals addressed within the scope of this project include entity recognizers development, profile definition and documents linkage through the use of transformers technologies. As a result, the relevant parts of the documents to be extracted are related not only to the title and afiliation of the authors, but also to article topics such as references, which are also considered relevant parts of the scientific article.
volumes of textual data in the form of digital documents of a scientific-technical nature. This facilitates systematic Health research organizations have always been an ex- analysis, environment assessment and tje proposition of ceptional environment for identifying specific needs and plausible future scenarios. This project significantly supgenerating new ideas that lead to innovative processes, ports the transition from a traditional model of medicine, products and services that improve health outcomes and known as reactive or curative medicine, in which patients the sustainability of health systems. Additionally, these go to the doctor and are treated, to a medicine that is organizations also provide key information on the en- not satisfied with curing, but seeks to prevent, improve vironment and market trends in their scientific areas of the quality of life, adapt to the individual, predict the interest. evolution of the disease and put the patient at the center.
Nevertheless, despite having technological surveil- In this process of change towards a new paradigm, lance (in many cases) or competitive intelligence systems as in the case of 5P medicine — which stands for more that enable them to configure customized alerts or the preventive, participatory, personalized, predictive and option of performing specialized information retrieval population-based medicine—, HLTs, and tools such as searches, they still lack solutions that support the sys- the one proposed in this project, play a fundamental tematic analysis of the large volume of information to role owing to the great potential ofered by health data retrieve the desired information. More specifically, ac- collection and the analysis of large collections of health cording to industry figures, medical professionals can data. spend up to 20 percent 1 of their working day conducting A convenient RDI planning, adapted to both the orinformation searches to support their daily activities. ganization and its environment, will guarantee the opti
In this context, in order to provide greater value, both mization and investment in technological developments to society and to the people who make up the structure that meet society’s demands. Thus, this would ensure of health research institutes, we propose the implemen- efectively translating these aspects into clinical practice tation of a platform for the advanced analysis of large through the productive ecosystem. Trend identification in research that consequently creates new markets will lfourish the development of companies which respond to these new business niches. This will translate into a significant improvement in patients’ quality of life and in the generation of wealth and well-being in society
The main objective of this project is the research and development of T2KNOW, an advanced text analysis platform based on Natural Language Processing (NLP) technologies, is the extraction and representation of semantic profiles of digital entities and the identification of research trends from the automatic analysis of scientifictechnical documents. Starting from this general objective, the project has the following specific objectives: • To design and develop a flexible, scalable and robust technological architecture for the management and processing of large volumes of unstructured data (text) as a necessary basis for advanced analysis. • To research and develop advanced text analysis algorithms, using PLN techniques, that allow knowledge extraction and semantic exploration of content for the detection of research trends. • To develop data visualization technologies to discover and graphically represent the evolution of research lines, topics and emerging technologies that allow the identification of research trends. • To design and execute a pilot test to validate the technologies developed in a key area such as healthcare, with the creation of specific corpus of scientific publications.
The process of knowledge discovery from natural language can be seen as a flow composed of several stages: from the initial text to a final relevant semantic knowledge representation in the form of an ontology. The ifrst step consists in the manual or semi-automatic construction of annotated linguistic resources. This requires choosing or defining an annotation scheme that is conducive to the domain of interest. From an annotated corpus, machine learning algorithms are trained to apply the same annotation scheme to large volumes of text. 3. Human language technologies Subsequently, all automatically discovered entities and relationships are grouped into a semantic graph. At this The T2Know project is a consortium involving several point, it is possible to perform post-processing tasks to entities such as ISABIAL (Institute of Health and Biomedeliminate redundancies, combine similar entities, or de- ical Research of Alicante), the company DIFUSION S.L. tect inconsistencies. Finally, a unified semantic structure and the University of Alicante. In addition, stakholders is obtained, which can be presented in the ontology for- such as AIMPLAS (Instituto Tecnológico del Plástico) and mat, where the relevant knowledge that was implicit in ITI (Centro Tecnológico de Investigación, Desarrollo e the original text is represented. Innovación TIC) were invited to be part of the team, since
To extract relevant knowledge from natural language they will contribute with expertise in areas related to the
text, PLN techniques have been introduced in systems health field such as plastics or technology. Depending on
such as ISODLE [
The use of natural language features can be used to modules, as shown in Figure 1.
build rule-based systems, such as the proposed OntoLT The focuses on three use cases or application areas
[
As we have seen, in the systems mentioned above, in order to extract knowledge from textual sources, it is necessary to contemplate the use and development of techniques for the semantic representation of knowledge, its storage and computational processing, and its metrics for the evaluation of its quality.
This task is responsible for semantically representing documents based on a series of characteristics previously extracted from them. This representation is governed by the scheme defined in the figure ??, which will allow to characterize the documents and in turn the entities and elements included in them, so that, with the use of semantic exploration techniques, it will be possible to recover not only documents but also their metadata as digital entities, with their respective profiles.
For instance, a given document may have a title, authors, an abstract, diferent topics and entities and also citations to other documents. In turn, authors may have an associated email, afiliation to one or several institutions, which may have an associated country. As can be seen, there are several characteristics that, when analyzed at a deeper level of detail, reveal the benefit of representing everything that can be characterized, since it enriches the quality of the metadata and thus ofers greater opportunities and points of view to explore the documents. Therefore, once documents are semantically represented, it will be possible to query not only documents, but also metadata under multiple non-conventional criteria such as, for example, Countries, institutions or authors most • Huntington – Epigenetics – Biomarker – Transcription – Next Generation Sequencing – Etc. • Multiple Sclerosis • Alzheimer – Myelin – Trained immunity – Olygodendrocyte – Epigenetics – Ageing – Inflammation – Etc. – Extracellular vesicles – Exosomes – Etc.
semi-supervised way by a human and using semantic techniques to discern what would be the appropriate link between profiles, either of documents or of any digital entity involved (ie. authors, institutions, countries, etc.). mation with semantic connotations are extracted from the textual content. These sensors must be able to identify, classify and extract all the necessary information to create a knowledge base. In this task, NERC domain and text classification sensors are developed, allowing the application of information and terminology extraction processes. In particular, the following phases or stages will be followed: • Terminology extraction. Term extraction based on statistical and linguistic algorithms. • Domain entity recognition. Development of entity recognizers specialized in each type of concept. • Information extraction through PLN tools, e.g., sensors of attributes defined in the document proifle such as date, author, language, summary, topics, entities, among others.
During this task, processes will be developed to link the document profiles through semantic relationships, as shown in the figure 3.3. That is, once the documents are semantically represented, it is necessary to link them to each other, and in this way they will serve as a connection point between other digital entities such as, for example, authors, subjects, etc. Not only will the documents serve as a link between digital entities, but also the common characteristics found between documents already semantically represented.
This linking of profile characteristics enables the inferring and discovering of new information otherwise harder to identify at first sight. For instance, authors or institutions that coincide on a particular topic or have written about a particular technology.
Profile linking presents the problem of semantic ambiguity [9] and ontology mapping [10] when it is proposed to automatically link profiles without running the risk of making mistakes. This issue can be dealt with in a
gories to which documents may belong, and to be able to classify within them the diferent types of domain 3.4. Data analytics and trends entities, it is necessary to reuse and develop diferent sensors based on natural language processing technologies. In this task, tools will be developed to detect the greatest Therefore, machine learning models will be developed, amount of statistical information related to the document adapted and reused for this purpose. profiles by temporal fractions. This will allow the trend
With these sensors it will be possible to classify docu- identification and topic evolution related to research in ments according to diferent topics, as well as to identify the sectors involved in the project. For this purpose, auand categorize diferent types of entities present in the tomatic learning techniques will be resorted to, such as contents, in order to guarantee an advanced exploration Time Series, as well as the potential ofered by their visuof the knowledge involved. To make possible the develop- alization. This visualization allows a two-fold result: (1) ment of these sensors based in part on machine learning, the generation of new knowledge, and (2) the presentawe will start from the corpus annotated by domain ex- tion of the temporal and statistical evolution of the pieces perts that will be used to train the PLN models. of information involved in the identified ontology. The processes involved in the ontology life cycle, 3.3. Profile linking namely, the ontology creation, management, analysis and reuse, entail workflows formed by several activities. These have been defined taking into consideration the main methodologies for the development of ontology models. Additionally, it is also necessary that these activities are supported by mechanisms and tools that allow their eficient development. These mechanisms are mainly robust visualization techniques and provided with an interaction that allows the user, through its capacity, to develop abstraction, conception, understanding, representation and learning of knowledge. One of the most important aspects to take into account in this task is semantic exploration and recommendation.
Given the existence of a semantic database, it is necessary to develop mechanisms for the exploration of the semantic network, supported by SPARQL(https://skos. um.es/TR/rdf-sparql-query/) or Cypher (https://neo4j. com/developer/cypher/) queries, of document profiles and other digital entities. These mechanisms will allow to retrieve not only documents through metadata filters, but also to make aggregate queries to discover statistical 13th international conference on World Wide Web, trends, and to make recommendations of profiles (e.g., ACM, 2004, pp. 100–110. documents, authors, institutions, topics, named entities, [6] E. Drymonas, K. Zervanou, E. G. Petrakis, Unsuetc.) through the semantic links that interconnect the pervised ontology acquisition from plain texts: the network. OntoGain system, in: International Conference on Application of Natural Language to Information Systems, Springer, 2010, pp. 277–287. 4. Conclusions [7] D. Faure, T. Poibeau, First experiments of using semantic knowledge learned by asium for informaCurrently, the project is in an initial stage focused on the tion extraction task using intex, in: Proceedings of capturing of both technical and functional requirements. the ECAI workshop on Ontology Learning, 2000. In addition, the KPIs, of key importance for identification, [8] T. Mitchell, W. Cohen, E. Hruschka, P. Talukdar, have been identified and the system development and B. Yang, J. Betteridge, A. Carlson, B. Dalvi, M. Gardevaluation will begin shortly. ner, B. Kisiel, et al., Never-ending learning, Communications of the ACM 61 (2018) 103–115.
Acknowledgments [9] Y. Gutiérrez, S. Vázquez, A. Montoyo, Spreading semantic information by word sense disambiguation, This project is funded by the Valencian Agency for Knowledge-Based Systems 132 (2017) 47–61. Innovation through the project INNEST/2022/24, par- [10] Y. Gutierrez, D. Tomas, I. Moreno, Developing an tially funded by the Generalitat Valenciana (Conselle- ontology schema for enriching and linking digital ria d’Educació, Investigació, Cultura i Esport) through media assets, Future Generation Computer Systems the following projects NL4DISMIS: TLHs for an Equal 101 (2019) 381–397. and Accessible Inclusive Society (CIPROM/2021/021) and T2Know: Platform for advanced analysis of scientifictechnical texts to extract trends and knowledge through NLP techniques.(Innest/2022/24). Moreover, it was backed by the work of two COST Actions: CA19134 “Distributed Knowledge Graphs” and CA19142 - “Leading Platform for European Citizens, Industries, Academia, and Policymakers in Media Accessibility”.