We believe that human language can be meaningfully mapped to a formal ontology for use in computational understanding of natural language expressions. We have created a formal ontology in first order logic (SUO-KIF: http://suo.ieee.org/suo-kif.html) called the Suggested Upper Merged Ontology (SUMO) http://ontology.teknowledge.com composed of roughly 1000 terms and several thousand formal statements about those terms. We have also created an index linking all 66,000 WordNet noun synset and 9000 verb synsets to terms in SUMO. We are working on using this index for natural language understanding tasks.
Our initial concentration has been on simple mappings, where there is a direct correspondence between an English noun and a class membership statement in logic, or where a verb can be mapped directly to a SUMO event type. For example, "The dog bites the man" shows both simple noun mappings and a simple verb mapping:
(exists (?D ?M ?E)
(and
(instance ?E Biting)
(instance ?D Canine)
(instance ?M MalePerson)
(agent ?E ?D)
(patient ?E ?M)))
The verb "bites" maps directly to the statement that there is an event of type "Biting". The noun "Dog" maps directly to the statement that there is an instance of the class "Canine" that participates in the event etc.
It is the case however that many possible sentences do not have this sort of simple mapping to concepts in an ontology. The next simplest case is where a word does not have a reified equivalent in the ontology, primarily due to the need for clarity and simplicity in the ontology, independent of the concerns of natural language understanding issues. Note that this is an important factor, since we intend that the ontology be appropriate for theorem proving tasks, and each of the terms in the ontology that are used to state the formal equivalent of English sentences must have an associated logical definition. Those terms and definitions must be stated in a manner that makes logical inference possible, and efficient.
A typical case of the lack of direct correspondence between English words and terms in the ontology is where the ontology models roles that agents play as relationships between the agent and a type of role, rather than reifying a subclass of agent filling the role. For example, "The pilot lands the plane." results in
(exists (?P ?PL ?E)
(and
(instance ?E Landing)
(instance ?P Person)
(attribute ?P Pilot)
(instance ?PL Airplane)
(agent ?E ?P)
(patient ?E ?PL)))
We are also dealing with more complex relationships where an entire phrase forms a pattern that has an equivalent template structure for a logical expression.
We are currently creating applications which use a restricted English language grammar as the input form for generating knowledge expressed in logic. Our expectation is that these applications will test our hypothesis that we can develop a grammar and translation rules that are powerful enough to allow humans to express most thoughts, while not encountering the overwhelming problems found in machine understanding of unrestricted human language.
| Workshop Homepage | Workshop Program | Previous Abstract | Next Abstract |