Martin Víta from Masaryk University, Brno, Czech Republic visited gave a talk on “Siamese Convolutional Neural Networks with Static Attention for Recognizing Faceted Entailment” at the ENGINE Centre on May 14, 2018. Martin was invited by Maciej Piasecki, the leader of G4.19 NLP group.

Recognizing textual entailment (RTE), i.e., a decision problem whether a sentence (called hypothesis) can be inferred from the given text, became a well established and widely studied task. As a consequence of the traditional binary or ternary class formulation, it is not possible to express the fact that a fragment of the hypothesis is entailed by the text, even though the “whole” entailment of the hypothesis from the text does not hold. The notions of partial textual entailment – and faceted entailment in particular – address this problem. Recently, there exist many approaches to RTE problem, but, in contrast, the task of recognizing faceted textual entailment is highly neglected. In this talk, we introduce a siamese convolutional neural network architecture with a simple static attention mechanism together with sentence compression and provide the first evaluations over modified SemEval 2013 Task 8 dataset.