Greenhouse Gas (GHG) accounting is traditionally a lengthy and manual process that requires the expertise of experienced environmental scientists; due to the recognition of the climate crisis through upcoming regulations on GHG accounting around the planet, the demand for tools that can support these environmental experts and accelerate their work is growing considerably at the moment. GHG accounting is merely one application of automated support tools that require the preservation of expert knowledge in a machine-readable and machine-understandable format; across fields, this is highly relevant for automating processes that today can only be performed by individuals with specialized training. In this paper, we present SAKE, a Semantic Authoring and Annotation tool for Knowledge Extraction that allows domain experts with no proficiency in semantic technologies annotating domainspecific PDF files, creating a Knowledge Graph with instances of standardized (or new) ontologies. The resulting Knowledge Graph can then be integrated into systems to automate specialized processes. SAKE has been developed together with domain experts in the field of environmental science and is currently used in the scope of a joint project on GHG accounting.
CEUR ceur-ws.org
From science to law and from research papers to regulatory documents, a large amount of textual
knowledge is today available in the form of PDF files. The knowledge transported through
these PDFs, while valuable for appropriately contextualized human readers, today remains
hard to integrate with automated systems. While current machine-learning methods, such as
large language models, mitigate this problem for content that aligns well with their training
data, these fall short for specialized knowledge that requires contextualized processing. Such
contextualization could be achieved if the information in a PDF was semantically integrated
with shared ontologies. This would not only enable automatic processing of the content, but
also—in-line with the core tenet of the Semantic Web—support the interlinking of pieces of
information across documents, institutions, and domains. While semantic annotation is readily
supported for HTML content, e.g., with Web-Annotation-based tools such as dokieli [
SAKE is a Web application built upon PDF.js2, a library developed by Mozilla that provides all the functionalities of a PDF reader. SAKE defines its own skin to ofer semantic annotation functionalities (see Figure 1(c)) next to the full capabilities of PDF.js. Specifically, SAKE enhances the PDF highlighting functionality to allow users to transform relevant content found in a document into structured knowledge expressed in the Resource Description Framework3 (RDF). SAKE’s current implementation uses AtomicData4 as a semantic back end. AtomicData hosts ontologies used for annotating documents and user data to add provenance information to annotations. In our implementation, AtomicData could be easily replaced by any other userbased graph database, such as Solid5 or GraphDB6.
To annotate a PDF file, a user (we consider domain experts) first loads an ontology (expressed in RDF) into SAKE’s semantic back end. The classes specified in this ontology are considered the user’s Known Concepts (KCs). SAKE displays all KCs on the right side of the user interface (see Figure 1). To annotate a PDF entity (text or figure), the user selects a KC and then selects the PDF entity. Then, SAKE displays a pop-up window that prompts the user for additional information 1https://wiser-climate.com/ 2https://mozilla.github.io/pdf.js/ 3https://www.w3.org/RDF/ 4https://atomicdata.dev/ 5https://solidproject.org/ 6https://graphdb.ontotext.com/ (see Figure 2) corresponding to the attributes and relationships related to the selected KC (i.e., object and data properties) and specified in the loaded ontology.
To ensure compatibility with all common PDF readers (e.g., Adobe Acrobat), the annotation
is stored as an RDFa string in the PDF document’s Content dictionary (cf. the PDF
specification [
Furthermore, SAKE integrates a Web server and responds to HTTP requests that specify appropriate content types (e.g., text/turtle or application/ld+json) with the graph embedded within the currently open PDF document. Finally, SAKE provides domain experts with the means to add new concepts and properties to existing ontologies; when a new concept is added, SAKE asks the expert to specify the corresponding HTML elements (e.g., a text field or a drop-down menu) to be displayed in the annotation pop-up window that is associated with the concept. This pop-up information is stored as a list of RDF instructions, which SAKE interprets at runtime. These instructions include validating and mapping strings to concepts in the ontology.
Today, GHG accounting is a time-consuming and expensive process that requires highly specialized environmental scientists to manually analyze companies’ processes, including their supply chains; even large multinational companies only commission these assessments rarely due to the amount of manual and expensive efort required. Faster and more cost-efective GHG assessment is required not only to comply with sustainability reporting obligations (e.g., the Swiss Ordinance on Climate Disclosures), but also to regularly assess current practices and reconsider company strategies to reach decarbonizations goals. In this context, WISER is an interdisciplinary project7 coordinated by Empa (Swiss Federal Laboratories for Materials Science and Technology) that aims at providing technological tools to increase the eficiency of GHG assessments. The project specifically required a way to capture knowledge from PDF documents as contextualized by the environmental scientists at Empa in a machine-understandable way. This applies primarily to Assessment Standards documents that must be followed when creating a GHG assessment and are published by diferent organizations (e.g., ISO, the European Commission, the World Business Council for Sustainable Development, or the World Resources Institute), which use idiosyncratic nomenclature and inconsistent concept definitions. Hence, two GHG assessment reports might not be comparable if diferent standards were followed or even if the same was followed but interpreted diferently.
To increase reproducibility and consistency across GHG assessment reports, WISER aims to create ontologies that describe diferent assessment standards and bridge ontologies that identify commonalities that permit the automatic translation of reports across assessment standards. Given that the environmental experts in our team are not ontologists, SAKE is proving value in capturing their knowledge when reading an assessment standard. The KG resulting from experts annotation will be incorporated in a Web application that accelerates the creation of GHG assessments and can translate reports from one assessment standard to another.
Providing non-semantic technologies experts with tools for creating semantically enriched
content has remained a challenge for several decades [
As described, most of the relevant related tools focus on HTML content, not on PDF documents as SAKE does). These documents hold vast amount of knowledge if read and annotated by experts. Moreover, SAKE is highly interested in high quality semantic annotations to integrate them in a tool (e.g., a dashboard) that can accelerate highly specialized real-world processes. To overcome one relevant entry barrier to using semantic technologies by domain experts, we have created SAKE, a tool that allows domain experts to create structured knowledge from PDF documents. This knowledge can then be exported as a KG and integrated into a tool for supporting highly specialized tasks such as GHG accounting. SAKE is provided with this publication as open source8, and remains in iterative development; it is currently used by environmental scientists in the scope of an interdisciplinary GHG accounting project. However, we expect the need for semantic annotation, sharing, and automated reasoning on top of extracted knowledge to keep growing across a variety of domains in which knowledge is still documented in PDFs, and their interpretations remain within the experts’ minds.
Acknowledgments: We thank Dr. Didier Beloin-Saint-Pierre, Alexander Kirsten, and Dr. Daniel Lachat, environmental scientists at Empa, for their support in testing SAKE. SAKE has been developed as part of the WISER flagship project funded by Innosuisse.