Home › Research Highlights

Interaction-situated Semantic Alignment

The OpenKnowledge project aims at lowering the cost of participation in semantic-intensive distributed systems by focusing on semantics related to interaction, which are acquired at low cost during participation, and using this to avoid dependency on a priori semantic agreements. We face the problem of semantic heterogeneity in multi-agent communication under this perspective, but we take a step further and rely only on interactions themselves to resolve terminological mismatches. We call this approach "Interaction-Situated Semantic Alignment" (or I-SSA, in short). I-SSA does not assume existence of any ontologies, neither local to interacting agents nor external to them. Our underlying claim is that semantic alignment is often relative to the particular interaction in which agents are engaged, and that, in such cases, the interaction state should be taken into account and brought into the alignment mechanism.

During the last few years, ontology matching has attracted the attention of researchers from different areas within Computer Science, mainly because it offers the possibility of avoiding the rigidity entailed by centralised ontologies. The state of the art shows that ontology matching is generally performed at design-time, prior to integration, which means, in our case, prior to agents entering an interaction. Furthermore, most current ontology matching techniques follow a classical functional approach, taking two or more ontologies as input and producing a semantic alignment of ontological entities as output. This involves several drawbacks. On the one hand, it limits the dynamism and openness. On the other, it keeps matching out of the context of the interaction. Semantic similarity of terms is established in an interaction-independent fashion, for example, by means of external sources such as WordNet, where relations as synonymy, among others, was determined prior to interaction and independently from it.

Recent approaches look at applying ontology matching at run-time and only among those fragments of the ontologies that are deemed relevant to the task at hand or to current interaction. This allows for openness and dynamism, and has the additional advantage that we do not need to access the entire ontologies (what is desirable when, for instance, ontologies constitute commercially confidential information). Now, despite these advantages, dynamic ontology matching techniques still follow a functional approach: when a mismatch occurs, semantic heterogeneity is solved applying current state-of-the-art ontology matching techniques, albeit only for a fragment and at run-time. Moreover, although done at run-time and more focused on relevant bits of the ontologies, matching is still done separately from the interaction.

But the meaning of certain terms are often very interaction-specific. For example, the semantic similarity that exists, in the context of an auction, between the Spanish term "remate" and the English expression "winning bid" is difficult to establish if we are left to rely solely on syntactic-based or structural matching techniques, or even on external sources such as dictionaries and thesauri. The term "remate" may have many different senses, and none of them may hint at its meaning as "winning bid". But it actually has this very precise meaning when uttered at a particular moment of the interaction happening during an auction.

Our approach attempts to overcome the mentioned drawbacks. I-SSA is a very parsimonious approach to the problem of semantic heterogeneity in multi-agent communication with the aim of complementing the previous solutions applied so far. Our claim is that semantic alignment is often relative to the particular interaction in which agents are engaged, and that, in such cases, the interaction state should be taken into account and brought into the alignment mechanism.

Related publications:

Egglue Powered