-In this paper we present J-ALINAs, a JADE-based Architecture for LINguistic Agents. The purpose of this architecture is to support communication between agents whose beliefs and intentions are driven by different, heterogeneous, knowledge models. This objective, often referred in literature as semantic coordination, can be carried on through the identification of specific agent roles and behaviors dedicated to the mediation of agents' knowledge. In particular, such minimal hypotheses suggest an intelligent exploitation of natural language based technologies (resources and systems) as a necessary choice for capturing those similarities between the different knowledge models of agents trying to communicate, which are not in any way formally ratified. We aim to provide a flexible framework to be adopted in open multi-agent environments across different scenarios, providing a further abstraction level from the underlying details related to specific semantic coordination approaches; a high-cohesion and low-coupling design, and an agent-interaction protocol make possible for the architecture to face non-ideal use cases optimizing the communications among the agents. We discuss significant design issues, provide a prototypical implementation based on the JADE platform and a case study MAPLE - integrating an ontology mapping component in the framework, showing flexibility of the architecture in real applications and its independence from any specific mapping algorithm. Finally we will look at the semantic coordination protocol we designed from a strictly formal perspective, providing a CCS (Milner's Calculus for Communicating Systems) description of the protocol itself.
Index Terms—Semantic coordination, ontology mapping, cooperating systems, intelligent agents.
Enabling communication between heterogeneous semantic peers (agents, semantic web services etc…) is a fundamental issue for the developing of the Semantic Web, especially in relation to the high levels of heterogeneity, evolution, distribution and autonomy of information which already characterize the Web as we know it now.
For the agents to be able to carry out the tasks they were designed for, they must impact the communication barrier between themselves: they will have to make their services available to the community, and recognize those of other actors in the community (to whichever extent it is considered, up to the Web as a whole); recognize, interpret and respond properly to communicative acts initiated by other agents and understand messages’ content.
Although an important effort has been done by FIPA (Foundation for Intelligent Physical Agents) towards a standardization of agent platforms and agent message transport, too many dimensions of variation still are to be taken into account.
First of all, while adherence to the FIPA message protocol grants agents with the ability to establish a communication, distinguish between requests and responses, and bind them to a given subject of conversation, nothing is said about the real content of these messages, thus making impossible for agents belonging to different communities to actually communicate. We still have to cope with the heterogeneities of agent societies and, mostly, with the diversities in agent frameworks design which are often developed for ad-hoc environments and applications. All these aspects represent a threat against largescale interaction among different multi-agent systems.
If peers were able to autonomously discovery each other’s services and to identify the specific problems which impede their communication, it would be much easier to devise solutions for supporting their communication, having a solid base to address the impedance mismatch among agents’ different knowledge representations from a shared starting point.
In a scenario like the one depicted above, it appears evident that agents cannot rely on any shared form of understanding, their inner knowledge, as well as the functionalities they expose, being expressed and modeled upon ruling principles which are not known a-priori.
Our approach aims to: - identify the fundamental actors and roles involved in the semantic coordination activity - define agent-based mediation paradigms, - design an agent-interaction protocol to support semantic coordination in a flexible fashion, and - model the semantic coordination process giving the designed framework a high abstraction level from the specific instance of the process itself.
Following these guidelines, we designed and developed J-ALINAs1, a (JADE based) Architecture for LINguistic Agents: an agent framework for supporting semantic coordination between heterogeneous semantic peers. Main feature of this architecture is the abstraction from the way actually semantic coordination is carried out by a specific instance of an agent system, still considering it the main shared goal for the community. While abstraction from the techniques is maintained, on the contrary, it is important to identify the different kind of resources which may play a role in the process, in order to make them available when their contributions is required. In particular, we consider as necessary an intelligent exploitation of natural language (from which the name J-ALINAs arises) resources and processors, as language is the sole form of shared knowledge which is inherently adopted when expressing (through label descriptors, concept documentation etc..) the formal content of knowledge models and resources.
Moreover, we achieved a clear distinction between decisionmaking and service-providing competences, realizing a crucial separation between specific problem-solving skills and strategic knowledge about their composition to address a complex target, shared among the system’s agents.
This allows to analyze new problems simply rearranging existing elements, and to improve or add functionalities without necessarily modifying the entire system.
A lot of work has been done by researchers towards flexible architectures to support semantic coordination in open multiagent systems.
In [
Within a multi-agent system however each agent expresses its own conceptualization, and in open systems interoperation between agents has to take into account their heterogeneous nature and background, that will likely take them not to fully understand each other, because of semantic misalignments which can easily arise.
A major requirement for agents interoperability [
To put it in Burstein’s and Uschold’s words [
Moreover, some interesting links between multi-agent
systems and grid computing are given in [
ALINAs [
This classification provides enough abstraction to ensure an incremental approach to systems development, whereas developers can focus on informative sources and resource agents and start to use available information before they design more advanced components in the system.
As stated in [
Linguistic enrichment of ontologies (this is the name we
gave to the process here described) requires a proper
exploitation of several linguistic resources, which, due to the
lack of defined standards for representation of linguistic
knowledge, often differentiate upon many aspects, like
structure, semantics, granularity of their content and
representation. This strong heterogeneity led us to the
development of the Linguistic Watermark [
Following the same intuition, it is also important, should an
ontology have been already linguistically enriched, to know in
advance to which extent and through which modalities this
enrichment process has been conducted. In our framework, this
information is represented by the Ontological Linguistic
Watermark [
MAPLE is a plug-in for the Ontology Editor Protégé [
It relies upon external linguistic resources – compliant to the Linguistic Watermark package – to obtain further information useful to find semantic correspondences between ontologies. The mapping process, before a deep inspection of the semantic characteristics of the two ontologies, preliminarily requires, at the linguistic level, to search for alternate expressions (synonyms) and/or glosses for the labels which describe ontology content, and also to identify proper translations in the more complex situation occurring when ontologies are labeled using different (natural) languages.
The generic meaning negotiation process between semantic peers usually has to start with a first hand-shake, in which information about the peers’ respective knowledge is exchanged (in the form of ontology namespaces, as an example). Should the two agents find that they are not able to communicate due to any form of incompatibility between their form of knowledge, they will first try to coordinate in order to invoke the figures which can support their communication.
Suppose an agent wants to query a search engine, to obtain a link to a remote document; so far the scenario is particularly known (Figure 2). We then introduce this situation in the Semantic Web context, whereas documents will be expressed against a formal description of the content they describe: an ontology.
The agent, to effectively understand the document’s content, will likely have the need to find mappings between its set of beliefs and conceptualization (its ontology), and the ontology the document is referred from.
To carry out the meaning negotiation and reach a knowledge model which is acceptable for both peers, the agent will request mapping service to a dedicated agent; the latter will take control of the whole negotiation process and will eventually need to request other agents’ support or information regarding involved ontologies.
However, it is particularly interesting, again, to notice how this scenario is independent from the mapping techniques the dedicated coordinating agent will implement, whether they be based upon finding lexical anchors upon mapping documents or aimed to spot schema-level similarities, or to perform a combination of both.
The JADE2 platform has been adopted to provide a prototypical implementation of the framework, in order to obtain a fully platform-independent prototype. JADE actually simplifies the development of agent platforms through a middleware complying to FIPA standards, implementing some of the required agents for the platform to be FIPA-compliant, and allowing agents to be movable from one machine to another.
J-ALINAs maintains much the same classification of agent
roles which has been presented in section III.B, and more in
details in [
For this reason we believe it will be – in certain circumstances – more correct referring to resource, service or control roles rather than agents, since an agent could embody more than one role in the mediation process.
In other words it will not necessarily exist a straight 1-to-1 relation between an agent and the role it assumes in the society, even though in the following paragraphs we will suppose it to exist, merely to have a distinct view of the identified competences. Here follows a more detailed description of the roles we have mentioned so far.
Linguistic agents encapsulate one or more linguistic resources, and provide an interface to access their content, through different kind of services.
The agent registers itself at a Directory Facilitator (DF) living in the agent platform, and publishes its services and the resources’ URIs.
An ontological agent is – together with the mapper agent – one of the fundamental actors of the process: usually it is such an agent to start the semantic coordination procedure and to request a conceptual mediation against another agent’s ontology. He shall then assume the requester role (with this name we will refer to the ontological agent requesting the meaning negotiation), and choose the responder (the ontological agent holding the destination ontology); after that he will query the DF to search agents providing mediation services. We will discuss the handshaking phase in details in section V.
For the semantic coordination to take part in an effective way, each of the ontological agents willing to communicate publish a fingerprint of the conceptualizations they encapsulate (the Ontological Linguistic Watermark, OLW), containing information about the linguistic enrichment processes (as described in section III.C) which contributed to describe their knowledge content. As the ontology may have not been subject to any linguistic enrichment process at all, the OLW may even be reduced to the sole information about the idiom used to represent its concept identifiers. By comparing ontologies’ watermarks, the mapper agent will then be able to easily decide which supporting agents are likely to be contacted, and negotiate the (natural) language(s) to be used throughout the whole mapping process.
The only agent in the society to have the needed intelligence and a thorough knowledge of the problem, of the actors and of the methodologies to be applied to carry on a meaning negotiation activity, is the mapper agent.
We first need to distinguish between two different kinds of mapper agents: those which provide semantic coordination facilities by inspecting available ontology mapping documents, that is, resources available on the web which state conceptual correspondences between ontology resources, and those which offer the same service by computing those correspondences on-the-fly.
The first agent will mostly carry on basic look-up over available mapping documents (even more than one, as agent ontologies may have been built compositionally from several available smaller ones), supported by inferential abilities to entail new mappings other than those which are explicitly declared.
For those cases where mapping documents are not available (or they do not cover completely the addressed conceptualizations) – presumably not a small fraction of real world situations – the second kind of mapping agent is invoked to try to individuate and establish possible semantic similarities by inspecting the structure of the two ontologies and by exploiting information provided by linguistic agents.
We designed a set of JADE behaviors, realizing the system’s reactive layers hierarchy. In the following paragraphs we will describe the most significant ones, moving our perspective from one agent to another.
The behavior encapsulating access to linguistic resources and providing linguistic support to the ontology mapping process is a cyclic one, enabling the linguistic agent to react to mapper agent’s stimuli. In other words it models a simple reactive behavior, implementing a reactive agent paradigm.
A linguistic agent in the setup procedure will add an instance of such behavior to its execution queue; it will eventually reply to queries originated by the mapper agent with semantic anchors for the particular term, or with a NOT_UNDERSTOOD FIPA message performative in such cases where the query is malformed (or, more simply, out of its comprehension).
Queries are matched by the agent against an applicationdefined message template, providing developers with high freedom of choice with respect to the type and number of linguistic services the agent publishes in the system.
The linguistic-querying task is delegated to a specific behavior residing in the mapper agent (Figure 3), which is responsible for persisting anchors upon the agent itself; however the concept of persisting anchors is absolutely abstract for the framework, and it is responsibility of the application developer to implement the persist and retrieval functions.
The provided ontological agent paradigm has been designed to act indifferently as requester or responder in the mapping process. We actually reckon this has positive implications whether towards bidirectional knowledge mediation between agents, and the integration of the application-specific ontological agent into the surrounding environment, reducing the number of different actors for the developers to take into account.
If the agent is required to act as requester in the process it will start discovering the surrounding environment, contacting agents and handshaking them.
Otherwise, if the stimulus is a PROPOSE FIPA message performative, the agent will eventually accept to participate in a mediation and wait for a mapper agent to ask for its ontological linguistic watermark or information about its ontology’s concepts and relations among them.
Concurrently, the ontological agent runs an instance of a simple reflex behaviour aimed to answer specific queries about ontology’s instances, properties or relations among concepts. This behaviour can be exploited by developers to fit messages exchange among agents to support the specific mapping algorithms, and does not affect the handshaking procedure at all.
Modules providing ontology mediation based upon external linguistic resources have been designed to give the handshaking procedure a well-known form, and at the same time to give developers using the framework the chance to easily extend the system with custom behaviours encapsulating application-specific mapping algorithms.
Indeed, after the handshaking phase took place the mapper agent shall own all the elements needed to conduct a generic ontology mediation, and to effectively decide which other agents in the society are likely to be involved in the process; in other words we provide developers embracing J-ALINAs architecture a solid and shared base to design their application on, building a further abstraction level from the underlying environment.
The abstraction itself comes to evidence even from a strictly sequential perspective: because every mapping process starts with a communication among agents and the system’s facilitator whereas the agents discovery the surrounding environment, we reckon it is convenient to let developers the chance not to care – to some extent – about the coordination and synchronization layer, focusing on their system’s peculiar aspects.
The interaction protocol for the semantic coordination is intended to be the whole project’s keystone, designed to adapt itself to different situations, minimizing the number of messages the agents must exchange – as it is normal to expect in a distributed environment – to achieve their communication’s main purpose: the mapping of agents’ ontologies (or, more specifically, of the concepts they need to express in their communication).
Throughout the protocol requisite elicitation and analysis phases, the chance for an agent to embody more than one role in the mediation process initially emerged as a problem: indeed, this implies the agent to be aware of its manifold functions while interacting, discovering and choosing which other agent to query. For instance, an agent implementing both the mapper and the ontological paradigm could prefer to rely upon its own mapping skills instead of contacting a third-party.
This brought the need for the protocol to adapt itself to variations in the system’s topology (Figure 4) and in the message flow among agents, and at the same time to ensure developers freedom of choice in deploying functionalities on application-specific agents.
After the setup phase, the agents are in a stand-by state. Whenever the requester receives the stimulus to initiate the process, an handshaking activity (Figure 5) introduces the communication. The DF will be queried to discover which other ontological agents are living on the platform. The requester agent will then elaborate the search results, and choose the passive agent (responder) in respect to fully application-defined criteria.
The mapping proposal is sent to the chosen responder agent, who will eventually accept the proposal, notifying the requester with an ACCEPT_PROPOSAL FIPA message performative including its ontology URI.
The requester will then query the DF asking for the presence of formal-model based mapper agents providing mappings between its ontology and the requester’s one.
Whether such an agent exists in the platform, the handshaking phase stops here, and comes the time for the requester to start querying the mapper with concept-mapping requests.
In the other case, another query is submitted to the DF, this time looking for the existence of a mapper agent based upon external linguistic resources exploitation. If the query is successful, the requester shall choose which mapper agent to contact and send out a PROPOSE message performative, waiting for the eventual acceptance from the counterpart. The proposal message will include the agent’s ontological linguistic watermark and the responder unique identifier.
The semantic coordination is now complete: the mapper agent has all the information needed to negotiate the natural language to use throughout the mediation, and choose which other agents to contact; the actual ontology mapping process can now start. Whilst no other agent was holding control of each others’ actions and of message flow in the system so far, now the decisional power goes completely to the mapper agent.
The adoption of such a protocol does not violate the dynamics
of the process and the heterogeneous nature of actors which
are typically associated to the idea of Semantic Coordination.
On the mere perspective of ontological agents (which could be
considered as the end users of Semantic Coordination), they
just need to be aware of the existence of Mapping Agents
willing to help them in the process of communicating with
other agents based on different kind of knowledge, and
implement (as they are expected to do in whatsoever scenario)
basic speech acts for requesting their help. Knowledge of an
ontological agent’s own linguistic expressivity is also
requested in our paradigm, but this in line with recent trends in
the Semantic Web area, where ontologies need to be expressed
in a linguistically motivated fashion [
As anticipated in III.D, MAPLE is a Protégé plug-in integrating ontology mapping functionalities in the Protégé ontology editing tool. MAPLE also provides ontology mapping as a standalone process, though it is not conceived for a distributed environment.
We succeeded in abstracting MAPLE’s algorithms and their relations with linguistic resources from its original scope, and provided an implementation for a mapping agent which fits in a distributed environment according to the described framework.
From MAPLE’s algorithms’ analysis what soon emerged is a linear relation between the cardinality of the target ontology and the number of queries the mapper agent is supposed to submit to linguistic resources (especially for the resource which is expressed in the natural language adopted for the communication). Standing on MAPLE mapping algorithms, not all the queries can be estimated in advance and sent as a unique request to a linguistic agent, thus requiring an heavy communication load between the mapper and the linguistic agents. This makes unacceptable the approach with autonomous linguistic agents to hold information sources, because the size of the involved ontologies is obviously not estimable in advance.
So we looked for a trade-off between the mapping process’ distribution and the communicational complexity among agents: this resulted in one single agent implementing both the mapper and the linguistic agent paradigms. This way we kept constant and independent from ontologies’ size the number of messages to be exchanged in the system, encapsulating in a single agent – thus in a single physical host – mapping algorithms and linguistic resources accesses.
The eventual need for a translation phase can easily be delegated to an external agent: the label(s) associated to the concept to map can be sent to a linguistic agent providing the translation services which will then send back to the mapped the translated terms; the mapper will then try to individuate the right translations among those provide by the translator, and then start to work in the language it is specialized in. This particular framework individuates a specific figure for mediation, given by a mapper agent which is particularly proficient in one (or more) idiom and which must be contacted only when its idiom pertains to at least one of the communicating ontological agents’ OLWs.
One of the most interesting perspective to look at JADE – and obviously at JADE-based systems – is the possibility to easily design – through extremely interesting extensions such as JADE-LEAP (Light Extensible Agent Platform) – agents living in mobile platform and devices.
Looking at the Semantic Web evolution together with the proliferation of wireless devices – PDAs and mobile phones – it becomes quite clear what the web is likely to become in the next years: a dense and highly dynamic network, populated by a multitude of nomadic elaboration nodes able to understand information sense, communicating in a more effective and intelligent way.
We believe it is natural – to some extent – to look at these nodes as autonomous entities communicating in a peer-to-peer fashion, living in an heterogeneous environment, providing and asking services one to each other; that is, actually, as agents.
As we already discussed, semantic coordination will be an essential and binding part of this communicative process: when most of the information will be expressed with respect to these specifications; knowledge mediation will be a very frequent runtime process, providing the base for mutual comprehension between systems.
Also, we believe a distributed approach has to be considered fundamental. This is one of the most interesting perspective to look at our work, which is actually providing a flexible framework for multi-agent systems development in response to the different needs coming from several backgrounds and practical applications.
We took into account the communicational effectiveness and efficiency, providing the framework with enough granularity to let developers easily control and keep constant the number of messages exchanged by the agents living in the particular instance of the system, leaving open the road to developments in scenarios with very specific constraints like those expressed in the mobile computing area.
In this paragraph we will give a – simple – formal
description of the handshaking phase and the coordination
protocol we designed, through Robin Milner’s Calculus for
Communicating Systems (CCS), [
We first define the generic Directory Facilitator as a recursive process:
DF = (search + searchFederated) . result . DF
We then define the twofold ontological agent behaviour as follows:
OA ¾R¾eq® OA
1 OA ¾P¾rop¾® OA
2 where Req stands for the receiving of a REQUEST FIPA performative, and Prop for the receiving of a PROPOSAL FIPA performative, and specify that:
OA1 = ( search .result.chooseP. proposeP . search .result). (chooseFM + search . chooseM . proposeWM ) where chooseP and proposeP indicate respectively the actions of choosing the passive ontological agent and notifying it with the proposal to collaborate in the mediation process, and chooseFM and chooseM indicate the action of choosing the
Formal Mapper agent, where available, and
OA2 = proposeP . watermark .
The mapper agent will behave as follows:
MA = proposeWM . storeWM . watermark . storeWM where WM stands for Watermark
Now we can define the whole handshaking process as processes above running concurrently and stimulating each other, limiting each process’ interface to allowed communication channels:
Handshake = (OA1 | DF | MA | OA2) \chooseP \chooseFM \chooseM \storeWM.\searchFederated