Repository for the paper Exploring Metaphoric Paraphrase Generation, accepted to CoNLL 2021.
Abstract: Metaphor generation is a difficult task, and has seen tremendous improvement with the advent of deep pretrained models. We focus here on the specific task of metaphoric paraphrase generation, in which we provide a literal sentence and generate a metaphoric sentence which paraphrases that input. We compare naive, "free" generation models with those that exploit forms of control over the generation process, adding additional information based on conceptual metaphor theory. We evaluate two methods for generating paired training data, which is then used to train T5 models for free and controlled generation. We use crowdsourcing to evaluate the results, showing that free models tend to generate more fluent paraphrases, while controlled models are better at generating novel metaphors. We then analyze evaluation metrics, showing that different metrics are necessary to capture different aspects of metaphoric paraphrasing. We release our data and models, as well as our annotated results in order to facilitate development of better evaluation metrics.
In this repository, you can find our code for compiling metaphor generation datasets,
both for controlled and free metaphor generation, leveraging MetaNet, FrameNet and BERT
(/metaphor-generation/build_training_data).
We use our novel dataset to fine-tune the T5 language model on controlled and
free metaphor generation (/metaphor-generation/fine_tuning/).
Finally, we evaluate our approach using both automatic and manual scores, which you can also find
in the repository (/metaphor-generation/fine_tuning/).
Some of the resources which we created for and during fine-tuning can be found in /resources/.
When using, please cite the following:
@inproceedings{stowe-etal-2021-exploring,
title = "Exploring Metaphoric Paraphrase Generation",
author = "Stowe, Kevin and
Beck, Nils and
Gurevych, Iryna",
booktitle = "Proceedings of the 25th Conference on Computational Natural Language Learning",
month = nov,
year = "2021",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.conll-1.26",
pages = "323--336",
abstract = "Metaphor generation is a difficult task, and has seen tremendous improvement with the advent of deep pretrained models. We focus here on the specific task of metaphoric paraphrase generation, in which we provide a literal sentence and generate a metaphoric sentence which paraphrases that input. We compare naive, {``}free{''} generation models with those that exploit forms of control over the generation process, adding additional information based on conceptual metaphor theory. We evaluate two methods for generating paired training data, which is then used to train T5 models for free and controlled generation. We use crowdsourcing to evaluate the results, showing that free models tend to generate more fluent paraphrases, while controlled models are better at generating novel metaphors. We then analyze evaluation metrics, showing that different metrics are necessary to capture different aspects of metaphoric paraphrasing. We release our data and models, as well as our annotated results in order to facilitate development of better evaluation metrics.",
}
@thesis{Beck2021,
author = "Nils Beck",
title = "Transfer Learning for Conceptual Metaphor Generation",
school = "Technische Universität Darmstadt",
address = "Darmstadt, Germany",
year = 2021,
month = jun,
type = "Bachelor's thesis"}
Contact person: Kevin Stowe, stowe@ukp.informatik.tu-darmstadt.de
https://www.ukp.tu-darmstadt.de/
This repository contains experimental software and is published for the sole purpose of giving additional background details on the respective publication.