It Depends

Context plays a crucial role for semantic similarity measurements. Depending on aspects such as location, time, task and user preferences, the respects that play a role for the comparison of concepts may vary. This variance should be reflected in a change of the similarity measurements’ results. Although this dependency has been observed in various human subject tests and is accounted for in recent similarity theories, there is no generic model of the characteristics of similarity measurement so far. The objective of this research is the formalization of such a model, which clarifies the exact systematics of the dependency of similarity on context. From a practical perspective, this model must allow for the determination of the most influential contextual aspects for similarity measurements in a concrete application. Introduction & Motivation. “Are canals more similar to rivers than to harbours?”—“Well, it depends.” The assessment of the similarity of concepts (and also instances) is heavily influenced by context. Accordingly, similarity measurements must take context into account to produce cognitively plausible results, instead of calculating generic similarity values that do not consider the respects of a comparison [6]. One might even go further and doubt the utility of context-free similarity measurements [2], since people’s similarity ratings in everyday situations are always influenced by their current context. The objective of this research is hence the development of a generic context model for semantic similarity measurement that supports the identification of the respects that influence similarity for a given task. Even though the influence of context is beyond dispute, the modelling of context for similarity measurement has been either disregarded, or handled as a mere add-on to the actual similarity theory so far. A more comprehensive utilization of contextual information in similarity measurement can help to produce results that are better tailored to the current situation of an individual. Moreover, context enables the disambiguation of terms such as “long” or “heavy” on a personalized basis: taking a user’s profile into account, which may, for example, contain information on previously taken hiking tours, such terms can be quantified for computation in a similarity measurement. Without contextual information, the required quantification is basically a wild guess.