Current engineering design processes oftentimes involve transferring information from manufacturer-provided data sheets into domainspecific design tools. While most data sheets are provided only as PDF files, this remains a tedious and manual task. This paper presents the Data Sheets Annotation Tool (DSAT), which assists engineers in gathering the information required in the design process. Using an OntologyBased Information Extraction (OBIE) method, the properties of components are extracted from data sheets and subsequently presented in an integrated, web-based interface. Engineers can now review and correct these automatic annotations, before exporting them for further use. In the demonstration, we employ a real-world use case rooted in modelbased space-system engineering. We show how the automated process can extract relevant component-attributes from technical data sheets and how users can redact the results. We further highlight the impact of content and quality of the underlying ontologies.
An important, recurring task in many engineering projects is to gather
information about components to be used. These are described by their physical
properties (e.g., spatial dimensions or mass) and the interfaces they provide or
require (e.g., propelling force or power consumption). The acquired descriptions
are subsequently fed into domain-specific design tools like Virtual Satellite [
The information required is barely available in machine-processable form, but is usually provided by PDF files (see Fig. 1 for examples). Engineers are required to obtain data sheets of interest, scan these files, and manually copy Copyright c 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). the important pieces of information to their respective design tools. Not only is this process tedious and repetitive, it is also error-prone and time consuming.
Existing information extraction approaches can alleviate this task, in
particular, Ontology-Based Information Extraction (OBIE) has proven itself a valuable
tool to convert text into machine-readable structures as witnessed by [
In the following we present the Data Sheets Annotation Tool (DSAT) [
DSAT is a web-based application which displays the data sheets and allows engineers to annotate attributes that will be used by different tools1. As shown in 1 A link to demo video:
https://zenodo.org/record/4034478/files/DSAT-screen-record.mp4
Fig. 2, it allows users to upload and select their data sheets (area I). The selected data sheet is displayed and attributes can be highlighted on the left panel (area II). The highlighted text can be categorized as a key, value, or remark. Engineers can add/edit/delete key-value pairs in this control form (area III). Users can also add custom text in a comment box. One set of a key, a value, a remark, and a comment is denoted as an annotation.
Manual highlighting is needed when there is no initial knowledge of data sheet’s domain. If a domain-specific ontology is available, DSAT calls upon a server-side information extraction pipeline to automatically detect key-value pairs in data sheets, and the results will be highlighted on the area II. However, these detected values should be reviewed and edited by domain experts before being used further. All annotations created, both manually and automatically, are summarized on the bottom right panel (area IV).
DSAT was evaluated by targeted users, who are involved in satellite development process. The evaluation result on DSAT was collected qualitatively. The overall feedback is that DSAT is intuitive to use. However, the evaluation also revealed some areas for further improvement like supporting a wider range of data sheet formats or the ability to assign a single area of text to multiple annotations.
The manual annotations collected during this evaluation were also used to evaluate the automatic extraction by OBIE. The envisioned workflow will be presented to a wider audience to collect additional feedback and adapt it to its users’ needs.
To harmonize across the heterogeneity of terms we want to search in the data
sheets, we use an Ontology-Based Information Extraction (OBIE) approach.
The accuracy of the auto-extracted result depends highly on the domain specific
ontology. For example, when processing a star sensor data sheet, an ontology
describing the specific attributes of a star sensor has to be provided. Our initial
ontologies can be found in [
To cope with this semantic challenge, we include external knowledge bases
such as Wikidata [
We use Wikidata to find entities corresponding to occurring terms. The entities returned from Wikidata contain semantic information, e.g. alternative label, description, superclass, similar entities, etc. This information will be inserted into the initial ontology, so that it can be used to extend the scope when searching for attributes in the data sheets.
In case of multiple entities, referring to different concepts, are returned from Wikidata, we use the context of data sheets and definitions from the lexical database WordNet to disambiguate between them. WordNet encapsulates the different meanings of a word in one synset each. These synsets are composed of definition, examples of usage, and relations to other synsets like synonyms or hyponyms. First, we collect all domain-representing keywords from data sheets and find the synsets that correspond to the domain. After collecting a set of domain-representing keywords from the corpus of datasheets, the corresponding synsets from WordNet are retrieved. If there are multiple synsets for some keywords, the most coherent subset with respect to a semantic relatedness measure is chosen footnote In our experiments we used Wu-Palmer Similarity, but others will be explored in the future. . These synsets now provide enough context to disambiguate between the candidate entities of Wikidata by comparing the respective textual descriptions. The selected entities are then selected for enriching the ontologies. 3
In this demonstration we presented DSAT, a tool to support engineers in
extracting technical information from PDF data sheets. However, similar
challenges as in the space domain also arise for example in patent analysis [
Currently, manual corrections by users only pertain to the specific data sheet they were made in. We plan to leverage this expert knowledge to further enhance 2 The extended ontologies used in this demonstration can be found here: https://zenodo.org/record/4034478/files/enriched_ontology.zip the ontology over time. While synonyms of existing concepts are easy to include, the situation gets more complex with hypernyms, hyponyms, or even terms that are unrelated to already known attributes. Especially for these cases we want to provide users a direct access to the ontology. Since most users have little experience with ontologies, the respective interface needs to be intuitive and prevent the users from introducing inconsistencies into the underlying ontologies.