What Do Neural Networks Actually Learn, When They Learn to Identify Idioms?
- Marco Silvio Giuseppe Senaldi (Scuola Normale Superiore of Pisa)
- Yuri Bizzoni (University of Gothenburg)
- Alessandro Lenci (University of Pisa)
Abstract
In this ablation study we observed whether the abstractness and ambiguity of idioms constitute key factors for a Neural Network when classifying idioms vs literals. For 174 Italian idioms and literals, we collected concreteness and ambiguity judgments and extracted Word2vec and fastText vectors from itWaC. The dataset was split into 5 random training and test sets. We trained a NN on the entire training sets, after removing the most concrete literals and most abstract idioms and after removing the most ambiguous idioms. F1 decreased considerably when flattening concreteness. The results were replicated on an English dataset from the COCA corpus.
Keywords: Idioms, Idiomatic Expressions, Neural Networks, Semantic Compositionality, Concreteness, Semantic Ambiguity, Ablation
How to Cite:
Senaldi, M., Bizzoni, Y. & Lenci, A., (2019) “What Do Neural Networks Actually Learn, When They Learn to Identify Idioms?”, Society for Computation in Linguistics 2(1), 310-313. doi: https://doi.org/10.7275/x015-az15
Downloads:
Download PDF