RHB+: A Type-Oriented ILP System Learning from Positive Data

Abstract

This paper presents the type-oriented relational l ea rner RH B+. Attaching type information to hypotheses is effective in avoiding overgeneralization as well as enhancing readability and comprehensibility. In many areas, such as NLP, type information is actually available, while negative examples are not . Unfortunately, learning performance is usually poor if types are attached when only positive examples are available. RH B + makes use of type information to efficiently compute informativity from positive exa mples only and to judge a stopping condition. The new technique of dynamic type restriction by positive exa mpl es lets covered positive examples decide the types appropriate for th e current clause. The current version of RH B+, written in the typed logic programming language LIFE, directly manipulates types as structured background knowledge when operations related to types are required. These features make RH B+ efficient and effective in attaching types selected from thousands of possible types. This leads to advantages over several previous learners, such as FOIL and PROGOL . Experimental results demonstrate RH B+ 's fine performance for both artificial and real data.