Knowledge Tagger performs Named Entity Resolution (NER) in texts using relevant domain ontologies and semantic data as background knowledge. Its distinguishing characteristic is its disambiguation-related customization capabilities as it allows users to define and apply custom disambiguation evidence models, based on their knowledge about the domain(s) and expected content of the texts to be analyzed. In this demo we explain the structure and content of such evidence models and we demonstrate how, given a concrete resolution scenario, one may use our system to define and apply them to texts pertaining to this scenario.
In this paper we demonstrate Knowledge Tagger1, a system that utilizes background
semantic information, typically in the form of Linked Data, to accurately determine
the intended meaning of detected semantic entity references within texts. The system
is based on a novel corresponding framework [
More specifically, through a structured semi-automatic process the framework
enables i) the exploitation of this a priori knowledge for the selection of the subset of
domain semantic information that is optimal for the disambiguation scenario at hand,
ii) the use of this subset for the generation of corresponding evidence and iii) the use
of this evidence for the disambiguation of entities within the scenario’s texts. As we
have already shown in [
Such models define for given ontology entities which other entities and to what extent should be used as evidence towards their correct meaning interpretation (see Table 1). Their construction depends on the characteristics of the domain and the texts. For example, assume we want to disambiguate location references within textual descriptions of military conflicts like the following: “Siege of Tripolitsa occured near Tripoli with Theodoros Kolokotronis being the leader of the Greeks against Turkey”. The nature of these texts allows us to expect to find in them, among others, military conflicts, locations where these conflicts took place and people and groups that participated in them. This in turn allows us to use these entities as evidence for disambiguating one another. For example, in the above text the term “Tripoli” is mentioned along with terms like “Siege of Tripolitsa” (a battle that took place in Tripoli, Greece) and “Theodoros Kolokotronis” (the commander of the Greeks in this siege). Thus, it is fair to assume that this term refers to the Greek town of Tripoli rather than, for example, to Tripoli of Libya. Generalizing this, we may define the location disambiguation evidence model of Table 2 where, for instance, a populated place can be disambiguated by the military conflicts that took place in it (row 1) and by the military persons that fought in conflicts that took place in it (row 3).
To define this model in the Knowledge Tagger demo we work as follows. First we press the “Create New Evidence Model” button to reveal the model creation form. Then we give a name for the new model (e.g. “Locations in Military Conflict Texts”) and we start filling the table form with the information of Table 2 (see Figure 1). First we select the target concept (e.g. “PopulatedPlace”), then the one to be used as evidence (e.g. “MilitaryConflict”) and then the (automatically calculated) relation path between them that we want to consider. For simplicity, in this demo we consider paths of maximum length two.
When the model is complete we press the “Generate Model” button to store the
model into the server and generate target-evidence entity pairs. Each pair is
accompanied by a degree that quantifies the evidential entity’s strength for the given target. (see
table 1). For example, James Montgomery acts as evidence for the disambiguation of
Beaufort County, South Carolina because he’s fought a battle there while his evidential
power for that location is 0.25, practically because there are 3 other military persons in
the ontology also named Montgomery. The exact way this strength is calculated may
be found in [
When the generation process is finished, the new model appears as an option in the list of defined evidence models and can be used to perform entity detection and disambiguation. To do that we select the model and then use the ”Input Text” form to perform NER to texts relevant to the scenario the model has been defined for. By pressing the “Perform NER” button the system works as follows: First it extracts from the text terms that possibly refer to the target entities as well as those that refer to their respective evidential entities. Then the disambiguation evidence model is used to compute for each extracted term the confidence that it refers to a particular target entity. The target entity with the highest confidence is expected to be the correct one. Figure 2 shows the results of executing this process on the above text about Siege of Tripolitsa.
Knowledge Tagger does not aim to be independent of the content or domain of the input texts but rather adaptable to them. That’s exactly its main differentiating feature from other similar systems as our purpose was not to build another generic disambiguation system but rather a reusable framework that can be adapted to the particular characteristics of the domain and application scenario at hand and exploit them to increase the task’s effectiveness.
The current version of the system’s user interface is still in an early stage of development. A first line of future work will focus on adding more domain knowledge to the system’s repository (other than the football and history datasets we already have) so that users are able to build evidence models for a larger range of domains. Moreover, we intend to allow users to use their own semantic data by linking our system to their repository.