|
---
|
|
license: cc-by-nc-sa-3.0
|
|
metrics:
|
|
- f1
|
|
- accuracy
|
|
widget:
|
|
- text: "We are at a relationship crossroad"
|
|
example_title: "Metaphoric1"
|
|
- text: "The car waits at a crossroad"
|
|
example_title: "Literal1"
|
|
- text: "I win the argument"
|
|
example_title: "Metaphoric2"
|
|
- text: "I win the game"
|
|
example_title: "Literal2"
|
|
---
|
|
|
|
# Multilingual-Metaphor-Detection
|
|
|
|
This page provides a fine-tuned multilingual language model [XLM-RoBERTa](https://arxiv.org/pdf/1911.02116.pdf) for metaphor detection on a token-level using the [Huggingface token-classification approach](https://huggingface.co/tasks/token-classification). Label 1 corresponds to metaphoric usage.
|
|
|
|
# Dataset
|
|
The dataset the model is trained on is the [VU Amsterdam Metaphor Corpus](http://www.vismet.org/metcor/documentation/home.html) that was annotated on a word-level following the metaphor identification protocol. The training corpus is restricted to English, however, XLM-R shows decent zero-shot performances when tested on other languages.
|
|
|
|
# Results
|
|
Following the evaluation criteria from the [2020 Second Shared Task on Metaphor detection](https://competitions.codalab.org/competitions/22188#results) our model achieves a F1-Score of 0.76 for the metaphor-class when training XLM-R<sub>Base</sub> and 0.77 when training XLM-R<sub>Large.</sub>.
|
|
|
|
We train for 8 epochs loading the model with the best evaluation performance at the end and using a learning rate of 2e-5. From the allocated training data 10% are utilized for validation while the final test set is being kept seperate and only used for the final evaluation.
|
|
|
|
# Code for Training
|
|
The training and evaluation code is available on [Github](https://github.com/lwachowiak/Multilingual-Metaphor-Detection/) |
