PIKES represents all the information contents in an RDF model organized in three distinct yet interlinked representation layers: Text layer, Mention layer, and Instance layer. The main classes and properties of the model are shown in the figure below, using an UML-like notation and different colors for the different layers (text in red, mentions in green, instances in blue).
This is the textual content from which knowledge is extracted. It consists of ks:Resources (i.e., documents) identified by URIs. Each resource consists of a raw text and an accompanying RDF description rooted at the resource URI, including metadata attributes such as dct:title and dct:created.
This layer consists of ks:Mentions. Different types of mentions are defined in PIKES model. ks:InstanceMentions denote instances of the domain of discourse and are further specialized based on the type of instance: ks:FrameMention for frame instances; ks:NameMention for named instances; ks:TimeMention for time intervals; and ks:AttributeMention for instances in the value space of lexical attributes (e.g., ‘very strong’). ks:ParticipationMentions link argument instances to participated frame instances (e.g., ‘fight of HIV’ links argument ‘HIV’ to frame ‘fight’). ks:CoreferenceMentions comprise spans of text having the same referent.
The instance layer describes the things of interest contained in a textual resource, abstracting from the actual ways they are expressed in the text. The main objects are ks:Instances of persons, organizations, locations, frames, dates and other entities of the domain of discourse. Instances are typed with respect to various taxonomies, are enriched with textual properties (e.g., rdfs:label and foaf:name), and are linked by a number of relations, including owl:sameAs assertions triggered by ks:CoreferenceMentions and frame-argument participation assertions triggered by ks:ParticipationMentions where the property conveys the role played by the argument. In this representation, frame instances reify complex relationships and are the main vehicle for relating instances.
Mention and Text layers are related by ks:mentionOf that links a ks:Mention to the ks:Resource it belongs to. Mention and Instance layers are related by three properties: ks:denotes, ks:implies, and ks:expresses:
- ks:denotes links a ks:InstanceMention to the ks:Instance it denotes.
- ks:implies links a ks:FrameMention to another instance (besides the denoted one) whose existence is implied by that mention, a situation occurring in case of argument nominalization.
- ks:expresses links a ks:Mention to the Instance layer assertions it expresses (i.e., that can be derived from it). Its RDF representation makes use of named graphs: each assertion of the Instance layer is placed in a named graph that represents the set of mentions (in some cases a single mention) that ks:expresses that particular assertion; ks:expresses is then asserted between each mention URI and the graph URI.
Put together, properties ks:denotes, ks:implies and ks:expresses allow any instance and assertion in the Instance layer to be always referred to the mention(s) from where it was derived, thus enabling a fine grained tracking of the specific piece of text from where a bit of knowledge was extracted.
PIKES and the KnowledgeStore
PIKES representation model is compliant with (and represents a specialization of) the KnowledgeStore data model (see KnowledgeStore core data model, meaning that the input and output consumed and produced by PIKES can be used to populate a KnowledgeStore instance, where all the content processed and produced can be accessed, navigated, and queried in an integrated fashion.