Fusing automatically extracted semantic annotations

One of the necessary preconditions of the Semantic Web initiative is the availability of semantic data. The Web already contains large amounts of information intended for human users. This information is mainly stored as hypertext, which must be semantically annotated to make it accessible for software agents. The amount of information on the Web makes it impossible to solve the annotation task manually. So the usage of automatic information extraction algorithms is essential. These algorithms use various NLP and machine learning techniques to extract information from text (Ciravegna 2003; Cimiano 2005). The information extracted from different sources must then be integrated in a knowledge base, so that it can be queried in a uniform way. This integration process is called knowledge fusion. Semantic annotations extracted automatically will inevitably contain defective aspects, which can cause problems during integration. These defective aspects include the following (based on (Appriou, Ayoun et al. 2001)): 1. Ambiguity. In general, information which can be interpreted in several distinct ways is considered ambiguous. It applies to the case when it is impossible to decide what real-world item the information refers to. For instance, when dealing with geographical information a document describing a country with the name " Korea " does not allow a software agent to judge automatically, whether it refers to North or South Korea. 2. Uncertainty. Uncertainty refers to the case when it is not possible to say definitely whether a particular Boolean statement is true or false. For instance, information can be biased depending on the source it comes from. Also there is a possibility of incorrect extraction by the extraction algorithm. For instance, the NationByNation.com website often contains values (like " Russia's unemployment rate is equal to 1.5% "), which are different from other sources. This information should have lower reliability. 3. Imprecision. Sometimes the content of the statement can be imprecise itself. For example, it is possible that a statement contains a rounded number instead of the precise one (e.g., " $1.2 million " and " $1212000 "). 4. Incompleteness. In most cases an information source does not contain full information about the real-world item it describes. Some properties may be missing from the description. For instance, an entity such as a country has many properties (economic, geographic, etc.). It is unlikely that all source documents will explicitly mention all of them. 5. Vagueness. Sometimes a predicate of a statement can be …