Efficient Knowledge Acquisition for Extracting Temporal Relations

Proceedings of the Australasian Language Technology Workshop 2005, pages 87–95, Sydney, Australia, December 2005. Efficient Knowledge Acquisition for Extracting Temporal Relations Son Bao Pham and Achim Hoffmann School of Computer Science and Engineering University of New South Wales, Australia {sonp,achim}@cse.unsw.edu.au Abstract We present KAFTIE – an incremental knowledge acquisition framework which utilizes expert knowl- edge to build high quality knowledge base annota- tors. Using KAFTIE, a knowledge base was built based on a small data set that outperforms machine learning algorithms trained on a much bigger data set for the task of recognizing temporal relations. In particular, this can be incorporated to bootstrap the process of labeling data for domains where anno- tated data is not available. Furthermore, we demon- strate how machine learning can be utilized to re- duce the knowledge acquisition effort. 1 Introduction Recent years have seen growing interests in tempo- ral processing for many practical NLP applications. For example, question answering tasks try to find when and how long an event occurs or what events occur after a particular event. Several works have addressed temporal process- ing: identification and normalization of time ex- pressions (Mani and Wilson, 2000), time stamping of event clauses (Filatova and Hovy, 2001), tempo- rally ordering of events (Mani et al., 2003), recog- nizing time-event relations in TimeML (Boguraev and Ando, 2005). At a higher level, these temporal expressions and their relations are essential for the task of reasoning about time, for example, to find contradictory information (Fikes et al., 2003). In this emerging domain, there is a clear lack of a large annotated corpus to build machine learning classifiers for detecting temporal relations. We pur- sued the idea that an incremental knowledge acqui- sition approach could be used to develop a knowl- edge base of rules that utilizes experts’ knowledge to overcome the paucity of annotated data. In fact, this approach could be combined nicely with the process of annotating data. When a new piece of data is annotated differently to what an existing KB proposes, the annotator specifies a justification for the decision in the form of a rule which is added to the knowledge base. Our experience shows that the time it takes to formulate a rule explaining why the data is annotated in a certain form is not much if the users have already spent time on deciding on the annotation. This is particularly true for complex tasks e.g. annotating relations between temporal ex- pressions where it is not obvious whether there is any relation between temporal expressions. Impor- tantly, rule formulation time does not depend on the size of the knowledge base. Our incremental knowledge acquisition frame- work is inspired by Ripple Down Rules (Compton and Jansen, 1990) which allows users to add rules to the knowledge base (KB) incrementally while au- tomatically ensuring that the knowledge base is al- ways consistent. A new rule added to the KB is only applicable to those cases where the current knowl- edge base did not perform satisfactorily according to …

Content truncated to 500 words