The annotation of the GENIA corpus, a set of biomedical articles, targets the classi cation of biological entities based on their association with a domain-tailored taxonomy of categories. By incorporating information extraction process on the corpus we have developed a knowledge base (KB) that includes a more comprehensive taxonomy of categories, relationships between biological entities, and a hierarchy of relationships. We present our experiences in exploring the expressiveness of OWL to accommodate our KB. OWL proves to be su cient to accommodate the extracted knowledge, however, it lacks expressive power when generalization of knowledge is needed. To this end we endorse recent endeavors in extending OWL with fuzzy description logic.
The GENIA corpus [
We propose an extension to the original GENIA ontology that not only makes the original ontology more comprehensive for reasoners but also accounts for additional information embedded in the GENIA corpus. We are particularly interested in possibilities of encoding this new knowledge in OWL, an ontology language that is becoming increasingly popular in both academic and commercial sectors. The proposed extension to the original ontology involves 1) enriching conceptually the structure of the original taxonomy of categories, 2) asserting the category membership of biological entities, 3) introducing binary relationships between biological entities, 4) building the hierarchy of relationships, and 5) connecting the ontology to an external, well-developed source of knowledge.
We employed an information extraction process on the corpus in order to evince the above mentioned features based on which the ontology is built. The process produced a set of biological entities, relationships between them, and their lexical decomposition. We also associated the biological entities with their descriptions using the UMLS Metathesaurus, a large vocabulary database that contains information about biomedical and health related concepts.
The rst version of the extended GENIA ontology in OWL 1 has already
been published [
The original GENIA taxonomy consists of only the declaration of classes and axioms about subclasses. The taxonomy \conceals" potentially useful pieces of information, such as: 1) there is a few default terminal classes that serve as placeholders for instances that do not belong to any of these classes' siblings, 2) only terminal classes have instances, and 3) an instance belongs to a single class.
The fact that the biological entities can directly belong to the terminal classes only can be embedded in the ontology by introducing covering axioms, such that C D1t: : :tDn, where D1; : : : ; Dn are subclasses of class C, i.e., if an individual is a member of class C it must also be a member of at least one of C's subclasses. We further narrow the \at least one" expression to \exactly one" by declaring that the sibling classes are disjoint, i.e., D1 u : : : u Dn v?. The disjointness of classes is assumed by the fact that each distinct biological entity that appears in the corpus is annotated to one and only one class. The last axiom also addresses the issue of the default classes.
Asserting the membership of individuals (biological entities) is straightforward. Each annotated (and preprocessed) biological entity is a member of a class indicated by the annotation.
Due to the OWL entity naming constraints the names of individuals are encoded (but still fully understandable by humans). For the sake of clarity each individual additionally carries a label property that contains the original form of the biological entity.
We also introduce the hasCUI property that links an individual with UMLS Metathesaurus through its Concept Unique Identi ers. Following nested tags in the corpus we add additional properties, stemsFrom and isRootFor, that hold between individuals, the name one of which is lexically composed from the other.
The identi cation of acronyms (during the information extraction process) leads to another fact that can be stated about two individuals, one of which is an acronym of the other, namely that the two are the same.
The TBox of our ontology is extended by declaring a set of object properties that will be used to assert verb relationships between biological entities in the corpus. These relationships are fed by the relationship extraction process. The extracted triples in the form of (subject, verb expression, object ) are used to enrich both the TBox and the ABox, by (1) deriving a hierarchical structure of object properties based on the syntax of verb expressions, and (2) asserting a relationship between the subject and the object, described by the verb expression.
The hierarchy of verb expressions is built by looking for expressions that have the same verb but di erent prepositions. We make an exception to this rule whenever the preposition by is encountered, which suggests that a verb expression with this preposition is in inverse relation to the verb alone.
The classes in our ontology mostly serve as \containers" for individuals, i.e., we cannot directly state anything about how they are related other than what was already discussed about their hierarchy. However, some information about classes can be inferred from the relationships between individuals, which is explicitly stated in the ontology by introducing additional axioms using existential and/or universal quanti cations. By using the universal quanti cation construct we state that for each class CD: CD v 8R:(CR1 t : : : t CRn), where CRi; : : : ; CRn is a set of ller classes. This set of classes is obtained directly from all the relationship triples with the object property R which have an individual of class CD on the left-hand side of the triple. The set of llers, i.e., right-hand side classes, is obtained by looking up the membership of the right-hand side individuals. In order to enforce the existence of a relationship, the last axiom can be replaced with CD v (9R:CR1 t : : : t 9R:CRn).
The two types of quanti cations can be combine to create a closure axiom of the form CD v (9R:CR1 t : : : t 9R:CRn) u 8R:(CR1 t : : : t CRn), or simply CD v 9R:> u 8R:(CR1 t : : : t CRn). The closure axioms suppress the open world assumption reasoning in OWL, i.e., they \close" the knowledge base to only what is known and derived from the corpus. 3
In the previous section we tried to infer some information about relationships
between classes. We achieved that by subsuming the classes with a set of
quanti cations, which resulted in extensive sets of axioms for each class. For some
applications, however, such super uous knowledge may actually be harmful. For
example, there has been a number of attempts (e.g., [
The process of generalizing relationships is usually based on assigning some con dence, with which a particular relationship holds between two classes, based on an individual membership distribution. By knowing the con dence of a relationship, useful information can be inferred about (1) the signi cance of such a relationship, and (2) the probability of an event that a pair of individuals will participate in this relationship.
Although the con dence of relationships can be calculated indirectly by an
application that works on an OWL ontology, this type of information cannot
be explicitly stated in OWL (in neither of its versions). OWL 2 provides the
expressiveness of the SROIQ DL language, and as such, does not allow for any
kind of uncertainty in individuals' membership, i.e., an individual (a pair of
individuals) either belongs to a concept (a relationship) or not. This uncertainty can
be introduced to the language by the fuzzy logic extension, originally proposed
by [
The importance of fuzzy description logic has already been recognized. Major
contributions in both de ning the problem and proposing the OWL syntax of
the fuzzy DL extension include [
The capability to embed uncertainty in asserting the membership of individuals would be useful to evince some kind of generalization and, to a certain degree, assess a knowledge base.
We would like to thank Dr. Inge Christiaens for biomedical consultation.