Towards Incremental LexicalAcquisition in

In this paper we present an implemented system for the incremental acquisition and revision of lexical entries in HPSG. As a byproduct of parsing, linguistic properties of unknown words are inferred from their sentential context. After parsing, the corresponding lexical entries are updated accordingly. A key idea of our approach is to employ a gradual, information-based concept of \unknownness", where lexical entries may contain unknown, i.e. revisable pieces of information. Re-visable information is further classiied into specializable and generalizable kinds. The overall approach is compatible with standard HPSG analyses and poses only a few extra demands on the grammar. These include marking revisable information with one of two distinguished formal types and specifying revisability clauses that provide access to the places of revisable information within the parse result. The update mechanism revises the corresponding places in the lexical feature structures ii the context actually provides new information. Revision translates to simple type intersection in the case of specializable information, whereas generalizable information requires type union. We demonstrate the inferential capacity of our system by illustrating in detail how verbal argument structure is incrementally acquired.