| Description |
Linguists and philosophers since Aristotle have attempted to reduce
natural language semantics in general, and the semantics of eventualities
in particular, to a "language of mind", expressed in terms of various
collections of underlying language-independent primitive concepts.
While such systems have proved insightful enough to suggest that such a
universal conceptual representation in in some sense psychologically
real, the primitive relations proposed, based on oppositions like
agent-patient, event-state, etc., have remained incompletely
convincing. All such primitives seem in practice to be somewhat
language-specific, and not to be fully interpretable in the absence of other
information, both from the rest of the predication including the
specific main verb and arguments involved, and from common-sense
inference about the real world that it denotes. The present paper
proposes that the primitive concepts of the language of mind are
"hidden", or latent, and must be discovered automatically by
detecting consistent patters of entailment in the vast amounts of
text that are made available by the internet using automatic syntactic
parsers and machine learning to mine a form- and language- independent
semantic representation language for natural language semantics. The
representations involved combine a distributional representation of
ambiguity with a language of logical form.
|
