Abstract

Recent interest in negation has resulted in a variety of different annotation schemes for different application tasks, several vetted in shared task competitions. Current negation detection systems are trained and tested for a specific application task within a particular domain. The availability of a robust, general negation detection module that can be added to any text processing pipeline is still missing. In this work we propose a linguistically motivated trigger and scope approach for negation detection in general. The system, NEGATOR, introduces two baseline modules: the scope module to identify the syntactic scope for different negation triggers and a variety of trigger lists evaluated for that purpose, ranging from minimal to extensive. The scope module consists of a set of specialized transformation rules that determine the scope of a negation trigger using dependency graphs from parser output. NEGATOR is evaluated on different corpora from different genres with different annotation schemes to establish general usefulness and robustness. The NEGATOR system also participated in two shared task competitions which address specific issues related to negation. Both these tasks presented an opportunity to demonstrate that the NEGATOR system can be easily adapted and extended to meet specific task requirements. The parallel, comparative evaluations suggest that NEGATOR is indeed a robust baseline system that is domain and task independent.