Hans Uszkoreit

DFKI and Saarland University

Syntax and Semantics in the Automatic Extraction of Relations from Texts

Information Extraction has become the cover term for a rather
successful application area in language technology. Since our
implemented grammar-based syntactic processing but especially our
knowledge-based semantic processing are not (yet) powerful enough in
coverage for true automatic language understanding, shallower methods
have been developed for the selective extraction of relevant pieces of
information from large volumes of texts. These pieces of information
may be topics, named entities, or relations. Whereas we could witness
strong progress in the detection of named entities, the recognition of
instances of relations, in particular of n-ary relations still
constitutes a major challenge.

In our terminology, relation extraction encompasses many types of
relations: binary relations, n-ary relations, events and arguments. In
our own work we have followed a systematic and analytic machine
learning approach that separates different knowledge sources and
different types of errors more than the majority of research actions
in the area.  In this way we could apply and investigate different
types of syntactic and lexical-semantic analysis. We can show how and
to what degree a successful approach needs to master known problems in
syntax, semantics and discourse analysis. Our specific way of
preserving, processing and sharing mentions, discovered syntactic
patterns and rules also lends itself to new ways of assuring
reproducibility of results and collaboration among groups.

In the presentation, I will demonstrate the general approach and the
role of lexical and grammatical information. I will report on
experiments with data-driven and grammar-driven parsers, among them
HPSG parsers. I will then explain our semantic approach and point out
open problems.