Ontological Analysis of FrameNet for Natural Language Reasoning EKATERINA OVCHINNIKOVA 1 & ALESSANDRO OLTRAMARI 2 & STEFANO BORGO 2 & LAURE VIEU 2,3 & THEODORE ALEXANDROV 4 ( 1 University of Osnabrück, 2 LOA-ISTC-CNR Trento, 3 IRIT-CNRS Toulouse, 4 University of Bremen) In recent years, the NLP research has shown that semantic knowledge plays an impor- tant role in solving tasks which require reasoning, such as question answering, infor- mation extraction etc. Much attention has been paid to the development of the lexical- semantic resources. Two of these resources, namely WordNet (http://wordnet.princeton.edu) and FrameNet (http://framenet.icsi.berkeley.edu), have been widely involved in various NLP systems. FrameNet (FN) has a shorter history in applications than WordNet, but lately more and more researchers demonstrate its po- tential to improve the quality of question answering (Shen & Lapata, 2007) and recog- nizing textual entailment (Burchardt et. al, 2007). However, the resource still has sev- eral considerable shortcomings. For example, in previous studies it was found that low coverage of the current version of FN makes its successful application to the real tex- tual data difficult, see (De Cao et. al, 2008). We want to make a further step in the di- rection of improving FrameNet for the goals of natural language reasoning and take a look on the conceptual structure of the resource. FrameNet is based on frame seman- tics (Ruppenhofer et. al, 2006). The lexical meaning of content words in FN is represented in terms of frames which describe prototypical situations spoken about in natural language. Every frame contains a set of frame elements corresponding to the participants of the described situation. Semantic relations, such as inheritance, causa- tion, precedence, are defined on frames. In this study we show that in addition to in- completeness FN suffers from conceptual inconsistency which can prevent appropriate inferences. In order to discover and classify conceptual problems in FN we investigate the FrameNet-Annotated corpus for Textual Entailment, FATE (Burchardt & Pennac- chiotti, 2008). Then we propose a methodology for improving the conceptual organi- zation of FN. The main issue we focus on in our study is axiomatization and restruc- turing of the frame relations. The proposed methodology is based on ontological anal- ysis which presupposes linking frames to categories in a formal ontology. The benefits of using axiomatized ontologies for constraining computational lexical resources have been demonstrated in the literature, see e.g. (Prévot et. al, 2009). In this study we con- centrate on improving FN reasoning capabilities by means of an axiomatized ontology, that is DOLCE (www.loa-cnr.it/DOLCE.html), an ontology which has been developed in order to capture the categories involved in natural language and common-sense be- liefs. For supporting ontological choices we apply different measures of similarity be- tween frames, see below.