This paper presents preliminary results on a framework for the representation and visualization of text as OWL ontologies under an open-domain paradigm, where no a priori schema for the facts to be extracted is available. The extracted ontology is visually represented as a specifically tailored node-link diagram. The applicability of the approach is demonstrated on a use case from the patent domain.
Extracting ontologies from text can significantly facilitate knowledge
integration and querying, through semantic alignment and mediation [
Within such context, two main challenges emerge: 1) to ensure the translation of textual input into well-formed ontologies that facilitate knowledge integration and querying in a schema-agnostic fashion; and 2) to provide the means for comprehensive visualizations that foster the understanding of the extracted knowledge, particularly at the factual level, in an intuitive manner that appeals adequately to users of diverse backgrounds and with varying levels of expertise.
These challenges are sharply manifested in the patent domain. The highly specialized and cross-domain terminology used in patent documents makes it very difficult, if not impractical, to rely on the availability of predefined schemata for the extraction of knowledge relevant to the task at hand. Moreover, the inherent complexity of patent documents render effective visualizations key tools for assisting experts in quickly grasping the main elements and their interactions.
In this paper, we present preliminary results on a framework for the representation and visualization of text as an OWL ontology under an open-domain paradigm, and illustrate its application with a use case from the patent domain. Abstracting from the specifics of the various semantic parsing methodologies, we describe an entity-relation-centric model for OWL-based text representation together with a graphical notation for its visualization as node-link diagram. 2
In accordance with the twofold goal of the proposed framework, related approaches to ontology extraction and visualization are discussed in the following. 2.1
Although ontology learning and population from text have been the subject of
arduous research [
Both result in representations that explicitly cater for n-ary relations, which represent a critical share of relations for effectively capturing the richness of textual contents. However, LODifier compromises ontology design with choices such as blank nodes, whereas FRED ensures high compliance with best practices, but the presented translations and heuristic rules are specifically tailored to DRSs. Instead, our goal is to provide a model for the generation of OWL representations from text that avoids commitments to specifics of the predicate-argument structures. 2.2
Many approaches to graphically represent ontologies have been proposed in the
last couple of years [
This is different in visualizations of RDF and Linked Data that are
typically more oriented towards the ABox. Examples include RDF Gravity [
isParticipantIn Subclassmof
Subpropertymof Aiming to abstract from predicate-argument specifics while assuring maximal interoperability within the Semantic Web and LOD context, we developed a minimal reference model for generating ontological text representations at a factual level.4 Hence, our goal is to provide core classes and properties for capturing the ways in which the extracted entities are interrelated, and that can be applied across domains, serving as anchors for attaching application-tailored class and property hierarchies.
A key design decision has been to model the extracted relations as classes
rather than properties. This is motivated by the saliency of n-ary relations in
textual resources, and the incurred loss of semantics when, instead of preserving
the n-ary dependencies, they are broken down into binary relations [
In accordance with the aforementioned principles, the model comprises the
following core classes: Entity subsumes the set of physical objects, processes, and
substances; Relation captures n-ary interrelations between entities; Attribute
encompasses characteristic aspects of an entity that cannot exist without it.
Alongside, a minimal set of upper-level object properties connect individuals of
the three classes: participant allows to link entities to the relations in which
4 Modalities, such as belief, causality, and entailment, are not considered as they can
be covered through specialized ontologies and knowledge patterns.
5 http://www.ontologydesignpatterns.org/ont/dul/DUL.owl
6 http://www.adampease.org/OP/SUMO.owl
they participate; actor and undergoer specialize participant in order to
discriminate between direct participants (“who?”, “what?”, etc.) and
complementary ones (e.g., the [pump]actor pumps [water]undergoer), while circumstantial
is a specialization used as a catch-all property for other types of participation;
hasAttr is used to associate entities to their attributes; hasPart is used to
capture mereological relations between entities; lastly, the datatype property
relevance allows to capture the relevance of the extracted entities to the
matter being considered. Figure 1 visualizes the core vocabulary using VOWL [
The extracted predicate-argument structures can then be translated into OWL representations, according to the following rules: – For each extracted entity, attribute, or relation, a named individual is generated; for co-referential entities, i.e., entities referring to the same real-world object, a single individual is introduced. – For each added named individual, respective rdf:type statements are added based on the extracted vocabulary of entities, attributes, and relations. – Respective rdfs:subClassOf axioms are added for each introduced entity, relation, and attribute class. – Instigative and passive participation links between entities and relations are translated into respective actor and undergoer property assertions; likewise for circumstantial participation, where additionally the prepositions lexicalizing the participation are defined as subproperties. For example, given the excerpt “...connected along...”, along is added as a subproperty of circumstantial. – Links between entities and attributes as well as entities and their parts are captured as hasAttr and hasPart property assertions, respectively.
The result is an OWL ontology consisting primarily of assertional knowledge, i.e., class and object property assertions, and to a lesser extent of terminological knowledge, as it could be derived from links to LOD resources, such as DBpedia and WordNet. Further specializations and schema enrichments, according to the given application needs, can be acquired through ontology learning. 4
Our visual notation for the graphical representation of the extracted ontology
is inspired by VOWL [
Entity Attribute
Patent documents are highly idiosyncratic, verbose texts that describe elaborate inventions and make heavy use of specialized terminology. These characteristics, in combination with the continuously growing rate at which patents are filed worldwide, incur extensive labor and time costs for carrying out typical patent portfolio analysis tasks. In this context, structured representations that can assist experts in identifying and contrasting patents relevant to the task in question, by rendering semantics explicit, and visualizations that effectively summarize the key elements of an invention and foster understanding, can entail immediate competitive advantages.
In the investigated use case, we address constructive patents, i.e., patents that describe the constituent parts of machine inventions and the ways in which they interact. In this context, it is important to specialize the described entities into components (e.g., coil, battery ), substances, processes, and other entities (e.g., temperature); likewise, for spatial and quantity attributes, such as inner charger pump pumps water shank along
extends along
drives extends in brushes plurality includes electric_toothbrush
comprises component substance spatial part other spatial quantity handle in is_mounted toothbrush_head
water_pas...
incorporates motor inner to along connects water_jet water_su... drive_shaft hasPart surface and plurality of brushes, as well as spatial parts (e.g., surface, bottom). To this end, the upper-level model definitions have been extended accordingly through the introduction of respective subclasses to the classes Entity and Attribute.
Using the mate tools [
Figure 3 shows the visualization resulting from the below patent claim, where the extracted entity, relation, and attribute individuals are outlined in respective fonts. The initial layout of the diagram has been generated with a force-directed algorithm and has then been manually adapted to increase its readability.
An electric toothbrush with a water jet, the toothbrush comprising a handle, a shank, and a toothbrush head that incorporates the water jet, in which an inner motor is mounted in the handle and the toothbrush includes a reciprocating drive shaft extending along the shank to drive a plurality of brushes in the brush head, including a water passage extending along the shank to connect a water supply to the water jet, and a pump in the shank for pumping water along the passage that is mechanically connected directly to the surface of the drive shaft.
In the given example, there are four types of extracted entities (components, substances, spatial parts, and other) and two types of attributes (spatial and quantity), as indicated by the different colors assigned to the entity and attribute nodes. As mentioned before, coreferential entities are captured by a single individual, upon which the respective participation links are projected. For example, the mentions of passage in “...a water passage extending along the shank...” and “...pumping water along the passage...” refer to the same passage entity; accordingly, there is a single “water passage” node to which the participation links in the extending and pumping relations have been projected.
All in all, the visualization provides an adequate representation of the patent claim that could be used to support analysts in understanding the elements and interrelations of the described invention. 6
In this paper, we have presented an upper-level model for extracting ontological text representations under an open-domain paradigm that allows abstracting from the specifics of predicate-argument structures, and a visual notation for its graphical representation that focuses on the visualization of facts rather than the ontological schema. The applicability of the proposed representation and visualization framework has been demonstrated through a use case from the patent domain.
Future work includes further validation and fine-tuning of the representation model, through extensive evaluation in cooperation with experts from the patent domain, as well as in an application-wise manner, where it will be used as the basis for assessing semantic similarity between patents. Furthermore, future research will have to address enhanced visualization paradigms that are more tailored to the patent domain.
General challenges with regard to the notation are improved scalability and readability of the visualization. A scalable visualization must be capable to represent larger ontologies extracted from several paragraphs of a text. In the patent use case, the individual claims could, for instance, form different subgraphs that are connected with each other according to specified dependencies.
Generalizing from the patent domain, the presented representation and visualization framework may serve as a valuable starting point for related cases of ontology extraction and visualization. The open-domain character of the ontology extraction and representation approach enables its wide application, along with the visual notation that combines the clarity of VOWL with an ABoxoriented view and capabilities to explicitly represent n-ary relations.
This work has been supported by the EU FP7-SME-606163 project iPatDoc.