The frame-based module of the Suiseki information extraction system

Christian Blaschke
CNB/CSIC

The Suiseki system for the extraction of interactions from large collections of
scientific text combines the statistical analysis of protein interactions, the
analysis of the syntactical structure of the phrase, and a frame-based module
dedicated to the detection of protein and gene names. The core of the system is
the set of frames that capture the different modes in which the relation between
proteins and genes are expressed in standard text. The approaches based on
frames are considered as a valuable alternative in specialized fields such the
one described here (Allen, 1995). We describe the details of the frames
currently included in the system, their intrinsic value, coverage and
performances. Even if at the detailed level the frame-based approach is able to
capture only a fraction of the interactions contained in different sentences
there exists a clear relation between the frequency of the detection of the
interactions and the accuracy of the information obtained. To the extent that
frequent interactions can be accurately detected with less than 20% error.
Therefore, we propose that the use of pre-defined frames in combination with
statistical and linguistical methods is a valid alternative for the analysis of
interaction networks described in the molecular biology literature.