Cross-lingual metaphor detection for low-resource languages

Abstract

State-of-the-art metaphor detection (MD) models achieve human-like performance for English data, while studies on MD for low-resource languages are currently missing. This thesis explores cross-lingual approaches that harness data from English, a high-resource language, in order to classify data in the target languages of Russian, German and Latin, either without using training data from the target languages or with as little as 20 instances. These instances were taken from the test data, but could also be created manually due to the small amount of annotating effort. The experiments indicate that the neural cross-lingual models mBERT (zero- and few-shot classification) and mBERT-based MAD-X perform well for German and Russian, while for languages where little data was used to pretrain mBERT, non-neural cross-lingual models with vector space model and conceptual features (abstractness, supersenses) outperform the mBERT-based models, if default hyperparameters are used. No validation data in the target languages was available for performing hyperparameter-tuning. Therefore, as a byproduct it was discovered that, while using a source language dataset for validation leads to overfitting, using a dataset from another language rather than the source language leads to decent results. This is especially true for the MAD-X model, which - with the help of successful hyperparameter-tuning - outperforms the non-neural classifier for the low-resource language Latin.