ICL-HD at SemEval-2016 Task 10: Improving the Detection of Minimal Semantic Units and their Meanings with an Ontology and Word Embeddings

Proceedings of SemEval-2016, pages 937–945, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics ICL-HD at SemEval-2016 Task 10: Improving the Detection of Minimal Semantic Units and their Meanings with an Ontology and Word Embeddings Angelika Kirilin and Felix Krauss and Yannick Versley Institute for Computational Linguistics Ruprecht-Karls-University Heidelberg {kirilin,krauss,versley}@cl.uni-heidelberg.de Abstract This paper presents our system submitted for SemEval 2016 Task 10: Detecting Minimal Semantic Units and their Meanings (DiM- SUM; Schneider, Hovy, et al., 2016). We extend AMALGrAM (Schneider and Smith, 2015) by tapping two additional information sources. The first information source uses a semantic knowledge base (YAGO3; Suchanek et al., 2007) to improve supersense tagging (SST) for named entities. The second in- formation source employs word embeddings (GloVe; Pennington et al., 2014) to cap- ture fine-grained latent semantics and there- fore improving the supersense identification for both nouns and verbs. We conduct a de- tailed evaluation and error analysis for our fea- tures and come to the conclusion that both our extensions lead to an improved detection for SST. 1 Introduction The SemEval 2016 Task 10 on Detecting Mini- mal Semantic Units of Meaning (DiMSUM) is con- cerned with the identification of semantic classes called supersenses for single words as well as multi- word expressions (MWEs). Identifying supersenses in text allows for abstrac- tions that characterize word meanings beyond su- perficial orthography (Schneider and Smith, 2015) as well as inferring representations that move to- wards language independence (Schneider, Mohit, et al., 2013). It has been used to extend named entity recognition (Ciaramita and Johnson, 2003) and to support supervised word sense disambiguation as it provides partial disambiguation (Ciaramita and Al- tun, 2006; Ciaramita and Johnson, 2003) as well as in syntactic parse re-ranking (Koo et al., 2005) as latent semantic features. The addition of MWEs - idiosyncratic interpreta- tions that cross word boundaries (Sag et al., 2002) - takes into account that the supersense of a MWE is usually not predictable from the meaning of the in- dividual lexemes. The Task moreover distinguishes between continuous MWEs like “high schooln.group” and discontinuous (gappy) MWEs like “track peo- ple downv.social”. The inventory of supersenses used for the task - 41 supersense classes, consisting of 26 noun and 15 verb supersenses - is derived from WordNet’s top-level hypernyms in the taxonomy. They are designed to be broad enough to encompass all nouns and verbs (Miller, 1990; Fellbaum, 1990). 2 Related Work Schneider and Smith (2015) are the first to approach SST and MWE detection jointly with a discrimina- tive model. Most of the previous work focuses on each of the tasks in separate. Sag et al. (2002) tried to raise attention for the issue of MWEs in general and analyzed different types of MWE. Baldwin et al. (2003) employed latent semantic analysis to deter- mine the decomposability of MWEs. Finally many MWE lexicons have been built for different pur- poses which is why Schneider et al. picked up the issue to address MWE annotation for general pur- poses (Schneider, …

Content truncated to 500 words