VENSES++: Adapting a deep semantic processing system to the identification of null instantiations

Proceedings of the 5th International Workshop on Semantic Evaluation, ACL 2010, pages 296–299, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics VENSES++: Adapting a deep semantic processing system to the identification of null instantiations Sara Tonelli Fondazione Bruno Kessler Trento, Italy. satonelli@fbk.eu Rodolfo Delmonte Università Ca’ Foscari Venezia, Italy. delmont@unive.it Abstract The system to spot INIs, DNIs and their anteced- ents is an adaptation of VENSES, a system for semantic evaluation that has been used for RTE challenges in the last 6 years. In the following we will briefly describe the system and then the ad- ditions we made to cope with the new task. In particular, we will discuss how we mapped the VENSES analysis to the representation of frame information in order to identify null instantia- tions in the text. 1 Introduction The SemEval-2010 task for linking events and their participants in discourse (Ruppenhofer et al., 2009) introduced a new issue w.r.t. the Se- mEval-2007 task “Frame Semantic Structure Ex- traction” (Baker et al., 2007), in that it focused on linking local semantic argument structures across sentence boundaries. Specifically, the task included first the identification of frames and frame elements in a text following the FrameNet paradigm (Baker et al., 1998), then the identifica- tion of locally uninstantiated roles (NIs). If these roles are indefinite (INI), they have to be marked as such and no antecedent has to be found. On the contrary, if they are definite (DNI), their coreferents have to be found in the wider dis- course context. The challenge comprised two tasks, namely the full task (semantic role recog- nition and labelling + NI linking) and the NIs only task, i.e. the identification of null instantia- tions and their referents given a test set with gold standard local semantic argument structure. We took part to the NIs only task by modify- ing the VENSES system for deep semantic pro- cessing and entailment recognition (Delmonte et al., 2005). In our approach, we assume that the identification of null instantiations is a complex task requiring different levels of semantic know- ledge and several processing steps. For this rea- son, we believe that the rich analysis performed by the pipeline architecture of VENSES is par- ticularly suitable for the task, also due to the small amount of training data available and the heterogeneity of NI phenomena. 2 The VENSES system VENSES is a reduced version of GETARUNS (Delmonte, 2008), a complete system for text understanding, whose backbone is LFG theory in its original version (Bresnan, 1982 and 2000). The system produces different levels of analysis, from syntax to discourse. However, three of them contribute most to the NI identification task: the lexico-semantic, the anaphora resolution and the deep semantic module. 2.1 The syntactic and lexico-semantic module The system produces a c(onstituent)-structure representation by means of a cascade of aug- mented FSA, then it uses this output to map lexi- cal information from a number of different lexica which however contain similar information re- lated to verb/adjective …

Content truncated to 500 words