Inducing Constraint-Based Grammars using a Domain Ontology

In many knowledge intensive applications, there is a critical need to populate knowledge bases rapidly and to keep them up to date. Since the World Wide Web is a large source of information that continuously is being updated, a solution is to automatically acquire knowledge from text, which requires language understanding in a smaller or grater degree. The need for “rapid” text-to-knowledge acquisition imposes some critical conditions on the methods used: scalability and adaptability. Thus, there is a need to move from handcrafted grammars and hand-build systems to learning methods. However, most statistical and learning techniques have been applied only to restricted domains (e.g., air travel domain) or tasks (e.g., information extraction where the knowledge is limited to a priori relations and entities), reducing the variety of the acquired knowledge. This thesis presents a framework for domain specific textto-knowledge acquisition, with focus on medical domain. The main challenge of this domain is the abundance of linguistic phenomena that require both syntactic and semantic information in order to “understand” the meaning of the text, and thus to acquire knowledge. Examples include prepositional phrases, coordinations, noun-noun compounds and nominalizations, phenomena which are not well covered by existing syntactic or semantic parsers. In my thesis, I propose a relational learning framework for the induction of a constraint-based grammar able to capture both syntax and aspects of meaning in an interleaved manner from a small number of semantically annotated data. The novelty of this framework is the learning method based on an ordered set of examples. This approach to learning follows the argument that language acquisition is an incremental process, in which simpler rules are acquired prior to complex ones. Several new theoretical concepts need to be tied together in order to make the approach feasible and theoretically sound: 1) a type of constraint-based grammar, called lexicalized well-founded grammar, which is learnable and able to capture large fragments of natural language; 2) a semantic representation, which we call semantic molecule that can be linked to the grammar and is simple enough to allow the relational learning of the grammar; 3) a small ordered set of semantically annotated examples, called representative examples, which is used as our training data; and 4) an ontology-based semantic interpretation encoded as a constraint at the grammar rule level (Φonto), which refrains from full logical analysis of meaning, known to be intractable. On the application side, the grammar learning is used for rapid acquisition of medical terminological knowledge from text. Semantic molecule. Given a natural language expression, w, we denote by w′ = h ./ b the semantic molecule of w, where h is the head acting as a valence for semantic composition, and b is the body acting as the semantic representation of w. The head is represented as a one level feature structure (i.e., feature values are atomic), while the body is a Canonical Logical Form given as a flat semantic representation, similar with Minimal Recursion Semantics (MRS). Unlike MRS, it uses as semantic primitives a set of framebased atomic predicates of the form: concept.attr=concept, suitable for the interpretation on the ontology: concept corresponds to a frame in the ontology and attr is a slot of the frame, encoding either a property or a relation. For example, for the adjective “chronic” we have the following semantic molecule:   cat a head X mod Y   ./ [X .isa = chronic,Y.Has prop = X ]