Scribd Logo
Upload

Welcome to Scribd!

  • 18 views

    A Knowledge Graph For Humanities Research

    Uploaded by

    Vayianos Pertsas
    The document presents a Knowledge Graph designed for researchers in the humanities that was created from over 25,000 papers from disciplines like archaeology and anthropology. The papers underwent natural language processing including removing noise from OCR text, classifying sentences, and extracting entities like activities, methods, and goals using deep learning models. Relationships between the extracted entities were also inferred to represent the research processes. The resulting Knowledge Graph contains on average 200 triples per paper and aims to help researchers gain insights by answering queries about problems, solutions, methods, and more across the literature.

    Copyright:

    © All Rights Reserved
    Download as DOCX, PDF, TXT or read online from Scribd

    A Knowledge Graph For Humanities Research

    Uploaded by

    Vayianos Pertsas
    18 views3 pages
    The document presents a Knowledge Graph designed for researchers in the humanities that was created from over 25,000 papers from disciplines like archaeology and anthropology. The papers underwent natural language processing including removing noise from OCR text, classifying sentences, and extracting entities like activities, methods, and goals using deep learning models. Relationships between the extracted entities were also inferred to represent the research processes. The resulting Knowledge Graph contains on average 200 triples per paper and aims to help researchers gain insights by answering queries about problems, solutions, methods, and more across the literature.

    Original Title

    A Knowledge Graph for Humanities Research

    Copyright

    © © All Rights Reserved

    Available Formats

    DOCX, PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    The document presents a Knowledge Graph designed for researchers in the humanities that was created from over 25,000 papers from disciplines like archaeology and anthropology. The papers underwent natural language processing including removing noise from OCR text, classifying sentences, and extracting entities like activities, methods, and goals using deep learning models. Relationships between the extracted entities were also inferred to represent the research processes. The resulting Knowledge Graph contains on average 200 triples per paper and aims to help researchers gain insights by answering queries about problems, solutions, methods, and more across the literature.

    Copyright:

    © All Rights Reserved
    Download as DOCX, PDF, TXT or read online from Scribd
    Download as docx, pdf, or txt
    18 views3 pages

    A Knowledge Graph For Humanities Research

    Uploaded by

    Vayianos Pertsas
    The document presents a Knowledge Graph designed for researchers in the humanities that was created from over 25,000 papers from disciplines like archaeology and anthropology. The papers underwent natural language processing including removing noise from OCR text, classifying sentences, and extracting entities like activities, methods, and goals using deep learning models. Relationships between the extracted entities were also inferred to represent the research processes. The resulting Knowledge Graph contains on average 200 triples per paper and aims to help researchers gain insights by answering queries about problems, solutions, methods, and more across the literature.

    Copyright:

    © All Rights Reserved
    Download as DOCX, PDF, TXT or read online from Scribd
    Download as docx, pdf, or txt
    You are on page 1of 3

    A Knowledge Graph for Humanities

    Research
    The steep increase of research publications in every major discipline makes it increasingly difficult
    for experts to maintain an overview of their domain, increases the risk of missing new work or
    reinventing solutions, and makes it harder to relate ideas from different domains. The latter
    becomes of high concern in multidisciplinary fields, like Digital Humanities, where maintaining a
    cross-disciplinary overview of what goes on in terms of research goals, activities and methods is
    even harder. This situation could be significantly alleviated by supporting information searches such
    as: find all papers that address a given problem; how was the problem solved; which methods are
    employed by whom in addressing particular research goals; etc. Answering queries like these
    essentially requires access to information that could be compiled interactively, or automatically
    extracted from research publications, finally offered in a structured form suitable for supporting
    semantic queries. Note that search engines widely used by researchers, such as Google Scholar1,
    Scopus2 or Semantic Scholar3 mostly leverage bibliographic metadata, while knowledge expressed in
    the actual text is exploited mostly by matching query terms to documents. In this paper we present a
    Knowledge Graph (KG) specifically designed for matching the above information needs of
    researchers in Humanities. The KG is derived from a large multidisciplinary dataset from the JSTOR 4
    repository, which has undergone various NLP processes so that information from each article could
    be properly extracted, combined with metadata and other information from the Web and finally
    transformed into RDF triples available as linked data. The entire process was driven by Scholarly
    Ontology (SO) [1], specifically designed for capturing research processes. A specialization of SO for
    DH is known as NeMO.

    The original dataset consisted of approximately 25,000 papers -produced by OCR on the original
    scanned files- from various disciplines such as Archeology, Paleontology, Social Sciences,
    Anthropology, etc. After a shallow rule-based cleaning to remove references sections and titles, the
    text was split into sentences using the SpaCy5 NLP framework. However, since the full text of each
    paper was the outcome of OCR, it had to be appropriately cleaned from noise elements such as
    unrecognized characters, tables, footnotes, references, section headings, etc. Furthermore, text
    deriving from scanned two-column papers yielded incomprehensible material that also had to be
    identified and removed. To this end we trained a Deep Learning text classifier using Hugging-Face 6
    BERT-base-uncased Transformer in order to recognize if a given sentence is proper or noisy, based
    on a manually curated dataset of 10,000 sentences (half of which were identified as clean and the
    rest as noise). Evaluation of our model yielded 95.8% F1 score for the text classification task. The
    classifier was then applied to the original dataset filtering approximately 3,700,000 cleaned
    sentences.

    Next, we trained three Deep Learning Entity Recognizers using Hugging-Face RoBERTa-base
    Transformers in order to identify and extract three core types of entities of Scholarly Ontology,
    namely: 1) Activities (i.e. actual research processes or steps thereof, like an archeological excavation,
    an anthropological study, an experiment, etc. carried out by the researchers-authors of the paper);
    2) Methods (i.e. procedures employed by researchers to carry out research activities, like an
    1
    https://scholar.google.com/
    2
    https://www.scopus.com/home.uri
    3
    https://www.semanticscholar.org/
    4
    https://www.jstor.org/
    5
    https://spacy.io/
    6
    https://huggingface.co/
    algorithm or a specific technique, which appear as named entities in text) and 3) Goals (i.e. the
    research tasks that were addressed by the researchers through their activities). Our training set
    consisted of 10,000 sentences, containing approximately 7200 Methods, 4200 Activities and 1800
    Goals. It was manually annotated by 3 annotators who, after appropriate training, reached inter-
    annotator agreement higher than 85% (Kappa statistic) for every task. Evaluation of our classifiers
    yielded F1 scores: 87.4% for Methods, 81.7% for Activities and 88.2% for Goals respectively. The
    classifiers were then applied to the original cleaned dataset.

    After entity extraction, additional post-processing rules were applied in order to infer semantic
    relationships among the extracted activities with methods and goals respectively. These rules are
    based on the SO definitions for employs(Activity,Method) and hasGoal(Activity,Goal) relationships
    and the proximity of their corresponding entities’ manifestations in text. Specifically, for each
    employs(Activity,Method) the corresponding textual spans of activity and method must overlap,
    while for hasGoal(Activity,Goal), co-appearance of the corresponding activity and goal in the same
    sentence is necessary. Evaluation of those rules on a test-set of 1000 cases for each relationship
    type, yielded F1 scores: 96.4% and 98.1% respectively.

    Finally, we extracted information from metadata regarding the authors of the articles (further
    matched, when possible, with ORCID7 using the provided API), publication information and author
    keywords. The KG was produced in RDF8 data format, using the NIF9 model for the URIs of the
    entities derived from text, which were then interrelated -when appropriate- and connected with
    those extracted from article’s metadata. Through this procedure, each paper is transformed into
    approximately 200 triples, on average.

    Generating KGs for scientific literature is an active research topic with many endeavors like [2,3],
    focusing on interconnecting bibliographic information of papers and authors, while others like [4,5]
    leveraging out-of-the-shelf NER solutions for extraction of named entities (e.g. material, task,
    dataset, etc.) in specialized domains of scientific literature. To the best of our knowledge, our KG is
    currently the only effort that concurrently addresses the problems of extracting information from
    articles’ full text, dealing with noisy OCRed material and semantically complex (and of variable
    length) entities like research activities and goals, while focusing on the domain of Humanities
    research. In addition, the inference of relationships between the extracted entities allows for better
    understanding and representation of their semantic context (e.g. the research process during which
    a method was employed, the reason for its employment, etc.) making it possible to address complex
    queries such us the ones described above. Future work involves expanding our KG with recognition
    and extraction of other SO entities such as researchers’ assertions based on the outcomes of their
    activities and information from citations, as well as linking our extracted entities with other linked
    data repositories and KGs across the web.

    References:
    1. Pertsas, V., et al. Scholarly Ontology: modelling scholarly practices, IJDL, Vol. 18(3), pp.173–190
    (2017).
    2. Steenwinckel, B. et al. Facilitating the Analysis of COVID-19 Literature Through a Knowledge
    Graph. ISWC 2020. https://doi.org/10.1007/978-3-030-62466-8_22
    3. Färber, Michael. The Microsoft Academic Knowledge Graph: A Linked Data Source with 8 Billion
    Triples of Scholarly Data. SEMWEB(2019).

    7
    https://orcid.org/
    8
    https://www.w3.org/RDF/
    9
    https://www.w3.org/community/bpmlod/wiki/NIF_Web_Services
    4. Dessì, D., et al. Generating knowledge graphs by employing Natural Language Processing and
    Machine Learning techniques within the scholarly domain. FGCS, 116 pp.253–264(2021).
    5. D’Souza, Jennifer and S. Auer. Computer Science Named Entity Recognition in the Open
    Research Knowledge Graph. ArXiv abs/2203.14579 (2022).

    You might also like