Illinois-LH: A Denotational and Distributional Approach to Semantics
Alice Lai, Julia Hockenmaier · 2014
This paper describes and analyzes our Se-mEval 2014 Task 1 system. Its features are based on distributional and denota-tional similarities; word alignment; nega-tion; and hypernym/hyponym, synonym, and antonym relations. 1 Task Description SemEval 2014 Task 1 (Marelli et al., 2014a) eval-uates system predictions of semantic relatedness (SR) and textual entailment (TE) relations on sen-tence pairs from the SICK dataset (Marelli et al., 2014b). The dataset is intended to test compo-sitional knowledge without requiring the world knowledge that is often required for paraphrase classification or Recognizing Textual Entailment tasks. SR scores range from 1 to 5. TE relations are ‘entailment, ’ ‘contradiction, ’ and ‘neutral.’ Our system uses features that depend on the amount of word overlap and alignment between the two sentences, the presence of negation, and the semantic similarities of the words and sub-strings that are not shared across the two sen-tences. We use simple distributional similarities as well as the recently proposed denotational sim-ilarities of Young et al. (2014), which are intended as more precise metrics for tasks that require en-tailment. Both similarity types are estimated on Young et al.’s corpus, which contains 31,783 im-ages of everyday scenes, each paired with five de-scriptive captions. 2 Our System Our system combines different sources of seman-tic similarity to predict semantic relatedness and This work is licensed under a Creative Commons At-tribution 4.0 International License. Page numbers and pro-ceedings footer are added by the organizers. License de-tails: