Metaphor Detection for Low Resource Languages: From Zero-Shot to Few-Shot Learning in Middle High German

Anonymous

Abstract: In this work, we present a novel unsupervisedmethod for adjective-noun metaphor detectionon low resource languages. We propose twonew approaches: First, a way of artificiallygenerating metaphor training examples andsecond, a novel way to find metaphors rely-ing only on word embeddings. The latter en-ables application for low resource languages.Our method is based on a transformation ofword embedding vectors into another vectorspace, in which the distance between the ad-jective word vector and the noun word vec-tor represents the metaphoricity of the wordpair. We train this method in a zero-shotpseudo-supervised manner by generating arti-ficial metaphor examples and show that ourapproach can be used to generate a metaphordataset with low annotation cost. It can thenbe used to finetune the system in a few-shotmanner. In our experiments we show the capa-bilities of the method in its unsupervised andin its supervised version. Additionally, we testit against a comparable unsupervised baselinemethod and a supervised variation of it.