The creation of schema.org as the de facto vocabulary for the implementation of Semantic Annotations was the dawn of a new era for the Web by motivating the Web developers to start weaving semantics in the content, mainly, for visibility reasons in search engine results. Moving further, the new version of the vocabulary enables Web entities to self-describe the Actions with which they interact with users, agents or services. In this scope, we present our ongoing work on automatic weaving of Actions based on the existing semantic annotations of a website.
The idea to build web agents to understand web content and interact with it in
order to realise a given plan and achieve goals has been part of the Semantic Web
vision [
In the same direction, but from another perspective, the discovery of Web
Services [
The remainder of the paper describes the structure of the actions dimension of schema.org and highlights the di culties of using it in section 2; outlines the ongoing work of the proposed approach in section 3; and closes with our conclusions and the current outlook in section 4.
One of the great features of the new extension is the exibility of combining any Action to any type of object/entity of the vocabulary. This is allowed by design as the related property has been hooked to the most generic class, the Thing. Speci cally, the action property is named potentialAction with domain the class Thing and range the class Action, which is at the top of the inheritance hierarchy of the actions. Thinking in triples, the schema has been designed in such a way that more than one triple could be created in order to enable a negrained description of the action, participating agents, objects, results of it and time variables.
For example, a Hotel object could participate in many di erent actions as shown in Fig. 1.
The above described exibility enables developers to build very descriptive websites and web applications. On the other hand, it needs deep understanding of the actions hierarchy for matching the two sides in order to make the right decisions that would result in meaningfully described content with a high potential of usability and visibility. Our ongoing work aims to tackle those obstacles in the scope of a broader idea of leveraging websites to open self-described APIs by annotating the content and the provided services. 3
The approach that we follow aims to address the di culties of developing Actions on websites and web applications. The added value of our contribution is the generation of suggestions about potential actions with which the described web entities of a website could support and interact. We de ne the problem as follows. We assume that semantic annotations have already been implemented for a given website, which contains entities that would be meaningful to interact with other entities in the Web sphere. For example, in case of lodging businesses, the o ered services (e.g. reserve rooms) could be potential objects that an agent would be interested to interact beyond a simple read-only consumption (e.g. to make/cancel or update a reservation). Therefore, let W = fE1; E2; : : : ; Eng with n 2 N be the set of all entities described in a website. Each entity E 2 W is the subject for properties such that E = PT [ PA, where PT stands for properties derived from any class and PA for those that refer to the action class. We make this arti cial distinction between properties in order to explain better our approach.
Furthermore, let T = fT1; T2; : : : ; Ttg with t 2 N be the set of all entity classes in schema.org and A = fA1; A2; : : : ; Amg with m 2 N be the set of all action classes in schema.org. We de ne the function : T ! P (A) that maps each entity to the set of actions that could be used by the entity type. Thus, having speci ed these mappings, we are able to de ne an entity E and its properties PT and PA. The core of our approach consists of the idea that we could de ne potential actions for a class T , by studying the range of the properties in an Action class A. In the scope of this paper, we identify two possible cases. The rst case, which is the simplest one, refers to the direct mapping of the range of PA to classes T . For example, according to Fig.1, the DepartAction A1 2 A has the property fromLocation PA1 2 PA with range Place T1 2 T . Thus, any subclass of T1 is in the range of PA1 and we could assume the A1 as a potential action for the class Hotel (Place ! Hotel).
In the second case, we assume that when properties PA of a class A1 2 A and properties PT of a class T1 2 T , and PT are not derived from superclasses (e.g. the Thing ), have matching ranges, then A1 is a potential action for T1. Building on the example of Fig.1 we apply the above idea as shown in Table 1. In particular, the LodgingReservation class and the ReserveAction of the schema.org vocabulary are described in the table with all their properties, excluding any derived properties from superclasses. The former models a reservation entity of a lodging business and the latter depicts an action of reserving concrete objects like a hotel or a restaurant. The middle column of the table indicates the elds that map from the Reservation class to the related Action class. As it is shown, the range of all the properties of the Action class map to a subset of the Reservation's properties.
At the evaluation part of the described approach, we employ datasets based on entities that instantiate schema.org classes and websites collected via web crawling in order to determine the occurrence of false positives and false negatives based on precision and recall measurements. An example of such dataset is our ongoing Linked Open Data dataset of touristic service providers in Innsbruck and its surroundings1. 4
Our approach could be summarised as an initiative towards the realisation of the paradigm that will transform websites to machine-readable data feeds, in other words, to APIs. At the core of this paradigm, lie the annotation of Web resources with a widely adopted vocabulary. The bene t of enabling the automatic extension of existing annotated entities with actions is threefold: a) supports the easy description of more content on a website, e.g. \Last reservation for the 24th of August was made 3 minutes ago"; b) promotes the exposure of speci c potential actions of existing entities of a Website, e.g. a potential action for the Hotel is a ReserveAction; and c) facilitates the establishment of a communication channel between agents and the provided services of a Website.
In the future work, we consider the EntryPoint property of the Action class,
which is specially interesting for Web agents, as it gives information about how
to formulate a request and what to expect as a response. Furthermore, we aim
to explore the matching of classes and actions with more ways and to integrate
it with the infrastructure provided by the Hydra [
Acknowledgments. This work has been partially supported by the EU projects BIG, BYTE, eFreight, FITMAN, LDBC, MSEE, PlanetData and PRELIDA, FWF project OntoHealth, as well as FFG projects OpenFridge and TourPack. 1 http://loi.sti2.at/openrdf-workbench/repositories/STI/summary