The SynDiKATe Text Knowledge Base Generator

The SynDiKA Te Text Knowledge Base Generator Udo Hahn Text Knowledge Engineering Lab Albert-Ludwigs-Universit¨at Freiburg D-79085 Freiburg, Germany hahn@coling.uni-freiburg.de Martin Romacker Text Knowledge Engineering Lab Albert-Ludwigs-Universit¨at Freiburg D-79085 Freiburg, Germany romac k er@coling.uni-freiburg.de ABSTRACT SynDiKA Te comprises a family of text understanding sys- tems for automatically acquiring kno wledge from real-w orld texts, viz. information tec hnology test rep orts and medical �nding rep orts. Their con ten t is transformed to formal rep- resen tation structures whic h constitute corresp onding text kno wledge bases. SynDiKA Te's arc hitecture in tegrates re- quiremen ts from the analysis of single sen tences, as w ell as those of referen tially link ed sen tences forming cohesiv e texts. Besides cen tering-based discourse analysis mec hanisms for pronominal, nominal and bridging anaphora, SynDiKA Te is supplied with a learning mo dule for automatically b o ot- strapping its domain kno wledge as text analysis pro ceeds. 1. INTRODUCTION The SynDiKA Te system b elongs to the broad family of information extraction (IE) systems [1]. Signi�can t progress has b een made already , as curren t IE systems pro vide robust shallo w text pro cessing suc h that frame-st yle templates are �lled with factual information ab out particular en tities (lo- cations, p ersons, ev en t t yp es, etc.) from the analyzed do c- umen ts. Nev ertheless, t ypical MUC-st yle systems are also limited in sev eral w a ys. They pro vide no inferencing ca- pabilities whic h allo w substan tial reasoning ab out the tem- plate �llers (hence, their understanding depth is lo w), and their p oten tial to deal with textual phenomena is highly constrained, if it is a v ailable at all. Also no v el and unex- p ected though p oten tially relev an t information whic h do es not matc h giv en template structures is hard to accoun t for, since system designers commit to a �xed collection of do- main kno wledge templates (i.e., they ha v e no concept learn- ing facilities). With SynDiKA Te, w e are addressing these shortcomings and aim at a more sophisticated lev el of kno wledge acqui- sition from real-w orld texts. The do cumen ts w e deal with are tec hnical narrativ es in German language tak en from t w o domains, viz. test rep orts from the information tec hnology (IT) domain as pro cessed b y the itSynDiKA Te system [8], . and �nding rep orts from a medical sub domain (MED), the framew ork of the medSynDiKA Te system [10, 9]. Our �rst goal is to extract conceptually and inferen tially ric her forms of kno wledge than those captured b y standard IE systems suc h as ev aluativ e assertions and comparisons [25, 24], tem- p oral [26] and spatial information [22]. Second, w e also w an t to dynamically enhance the set of kno wledge templates …

Content truncated to 500 words