At the ART group in Tor Vergata we have started our specific research on Textual Entailment in 2004, when the first Pascal Challenge on Recognizing Textual Entailment (RTE) has been launched.
Anyway, our NLP group has a long experience on lexical semantic, syntactic analysis and information extraction. This background helps us everyday in coping with such an intricate and complex problem. Our in-house linguistic systems and technologies give us the chance not only to start-up and improve methods for entailment detection and generation, but also to investigate better the linguistic and ontological nature of the phenomenon.
We took part at the RTE challenge with a prototypical system: please refer to the Challenge web page for more detail. At the moment, our research focuses on three main directions:

  • Entailment pattern acquisition: this research area aims at collecting generalized forms of Textual Entailment (such as X acquired Y entails X own Y), carefully inspecting textual corpora, using different techniques ranging from statistical counts to linguistic analysis. The goal is thus to create automatic methods able to build patterns and to apply them to texts to extract useful information to be applied to NLP application, such as QA and IE. In this view we are focusing on verb-based patterns, as verbs usually govern the meaning of linguistic relations.

  • Entailment recognition: the aim is to verify if an entailment relation holds (possibly with a degree of confidence) between two linguistic expressions. Unlike patterns acquisition, in the recognition task the textual material of T and H is given. The use of linguistic resources and analysis can be thus preferred. In this area, that has been the main focus of the RTE challenge, we are studying algorithm and strategies, based on different levels of knowledge, able to recognize entailment between sentential expression using graph distance measures applied to the syntatic graph produced by our in-house Choas syntactic parser. We are also investigating Machine Learning techniques able to detect entailment using specific and well-defined features: in this direction the main goal is to select the best suited features able to capture the relation, and the optimization of existing learning strategies (e.g., SVM).

  • Linguistic and cognitive investigation: as it is a complex and broad phenomenon, Textual Entailment cannot be only investigated by NLP techniques. A deeper linguistic and cognitive analysis of the problem is needed: linguistic, psyco-linguistic and cognitive aspects of entailment should suggest new directions and ideas for algorithms and resource use, as it has always been in AI. We are then trying to investigate more theoretical aspect of Textual Entailment, also with the help of experts in linguistics and cognitive science.
For any further detail please look at our publications.