Towards a Semantic Model for Textual Entailment Annotation

Introduction Annie Zaenen, Cleo Condoravdi, Valeria de Paiva The introduction of the rte (Recognizing Textual Entailment) para- digm (ref) in 2004 constituted a turning point in computational work on inference in natural language. That paradigm sees the task of de- termining the relation between two texts as one where assuming the truth of the first text, T, the thesis, leads to concluding to the (likely) truth or falsity of the other text, called H, the hypothesis. Later vari- ants extended the relation also to contradictions. It has focused work on textual inference on tasks that are on the one hand feasible and on the other hand based on real texts. In this volume we present some of the work that arose from this conceptualization of the task, mostly but not only focussing on methods that involve logical formalizations. The volume is based on three workshops that we organized in 2011 and 2012: the CSLI Workshop on Natural Logic, Proof Theory, and Com- putational Semantics on April 8 and 9, 2011 at Stanford,1 Semantics for textual inference on July 9/10, 2011 at the University of Colorado at Boulder,2 and the CSLI workshop on Semantic Representations for Textual Inference on March 9 and 10, 2012 at Stanford.3 The first paper in this volume, “The BIUTEE Research Platform for Transformation-based Textual Entailment Recognition” by Asher Stern and Ido Dagan, is from the lab that introduced the current textual inference paradigm, rte. The paper is an interim report on biutee, an open source and open architecture platform that allows users to develop components for textual inference systems. The sys- tem consists of a preprocessing module that does parsing, named en- 1http://www.stanford.edu/~icard/logic\&language/ 2http://www.stanford.edu/~cleoc/Sem-Text-Inf/ 3http://semanticrepresentation.stanford.edu ix LiLT Volume 9 Perspectives on Semantic Representations for Textual Inference. Copyright c ⃝ 2014, CSLI Publications. x / Annie Zaenen, Cleo Condoravdi, Valeria de Paiva tity recognition and coreference resolution and several modules that do inference recognition. The architecture allows the user to insert di↵er- ent lexical/knowledge resources (e.g. dirt, WordNet, Wikipedia) and to add his/her own modules. The inference recognition engine itself is transformation-based, constructing derivations that go from text to hypothesis via rewrite rules. This procedure that can be seen as an instance of a proof, with the di↵erence from traditional formal proofs being that the procedure allows for conclusions to likely true (false) instead of aiming for certainty. The system presented in this paper has meanwhile been integrated in an wider platform, excitement, that allows more flexibility.4 The hope is that ultimately the community interested in textual inference will have a platform comparable to that provided to the (statistical) translation community by moses.5 There are two other papers describing full systems. They concentrate on building systems that allow true textual entailments and are not geared to allowing conclusions that are only likely to be drawn by native speakers. The first paper “Is there a place for logic in recognizing textual en- tailment” by Johan Bos describes Nutcracker, a system that integrates Combinatory Categorial Grammar, Discourse Representation Theory …

Content truncated to 500 words