Bootstrapping an Ontology-based Information Extraction System Alexander Maedche, Günter Neumann, Steffen Staab (presented by D. Lonsdale) CS 652 – June 7/04
Traditional IE + machine learning Extensive use of NLP (SMES: German, English, Japanese) Ontologies and related tools (OntoEdit, OntoBroker) abstract ontology + lexicon concrete ontology Conclusions/reflections Overview
The mantra Lexical knowledge As usual, concepts are grounded in lexical items Extraction rules OntoBroker: deductive, OODB, F-Logic Ontology Abstract ontology + lexicon concrete ontology
Lexical knowledge Low-level lexicons, dynamically updated Basic low-level NLP: tokenization (50 classes) morphological processing POS tagging named entity extraction chunk parsing thematic role assignment (grammatical function) Cascading finite-state transducers
The NLP component
NLP terms Dependency syntax Chunk parsing Subcategorization Case Topolological fields PP attachment
Dependency syntax
Extraction Concept definitions Inference rules/axioms Bridging (forward inferencing) Syntactic dependency relations “...implementations of idiosyncratic syntactic cues for particular ontological structures...” Logical relations (e.g. transitivity, LocatedIn) OntoBroker engine
OntoEdit display (tourism)
An abstract ontology
A(n ontology) lexicon
Ontology learning So how does ontology learning happen? Ontology engineer specifies, refines knowledge structures Select and process a text corpus with the model Use a set of different learning approaches “...generalized association rule learning algorithm...” Extend the extracted model (all three parts...) Human reviews learning decisions The ontology is concrete, the methodology description less so...
The overall approach/system
GETESS visualization
Conclusions/reflections Heavy use of NLP (good/bad) Fairly typical mapping of lexical items, concepts, relations Toolkit approach: lingware, inferencing, GUI’s Machine learning description is vague A picture is only worth a thousand words...