The proposed algorithm required two training steps. First, a text classifier learns to classify metaphorical from literal sentences. 

Second, metaphorical sentences, which were correctly classified by our classifier as metaphorical, were passed to a reconstruction step, where metaphorical tokens were masked and then recovered through extraction (T5, BART) and masked language modeling (BERT, XLM-R).

Pipeline

Generation of new metaphors requires sentences from a corpus $$C$$ which $$MC$$ has classified as literal. A random token is masked and the masked metaphor model replaces the masked token, effectively trying to create a metaphorical text. Finally, use $$MC$$ to keep only truly transferred metaphorical texts.

University of Michigan - Ann Arbor

Google

This study presents a new approach to metaphorical paraphrase generation by masking literal tokens of literal sentences and unmasking them with metaphorical language models.

With this work we present a new metaphorical language generation perspective by transferring literal to metaphorical texts. The transfer concerns the replacement of tokens of different parts of speech, not only the common **verb** type Stowe et al. [2021], Chakrabarty et al. [2021], Yu and Wan [2019], but also **nouns** and **adjectives** (examples shown in Fig. 1).

Metaphorical Paraphrase Generation: Feeding Metaphorical Language Models with Literal Texts

Metaphorical text classification is a binary classification task trained on $I^{M|L}$.

Naive Bayes Rish, Random Forests, KNN, SVM, Logistic Regression, MLP, BERT, and XLM-R

Metaphorical text classification

Masked metaphor modeling is a task similar to masked language modeling, but the masked tokens are tagged metaphors. Fine-tuning a pre-trained model requires a limited number of texts classified as metaphorical (by $$MC$$) and with the metaphor tagged within the text.

Masked Metaphor Modeling

Generation of new metaphors

The inputs required for the metaphorical paraphrase generation approach include a corpus $$C$$ of unlabelled texts, a dataset $$I^{M|L}$$ of texts labelled as literal or metaphorical, and a classification threshold $$h$$.

Learn Before

Related

Learn After