Exploring Metaphoric Paraphrase Generation

Proceedings of the 25th Conference on Computational Natural Language Learning (CoNLL), pages 323–336 November 10–11, 2021. ©2021 Association for Computational Linguistics 323 Exploring Metaphoric Paraphrase Generation Kevin Stowe, Nils Beck, Iryna Gurevych Ubiquitous Knowledge Processing Lab, Technical University of Darmstadt https://www.informatik.tu-darmstadt.de/ukp/ Abstract Metaphor generation is a difficult task, and has seen tremendous improvement with the advent of deep pretrained models. We fo- cus 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 gen- eration process, adding additional information based on conceptual metaphor theory. We eval- uate two methods for generating paired train- ing data, which is then used to train T5 mod- els 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 bet- ter at generating novel metaphors. We then analyze evaluation metrics, showing that dif- ferent metrics are necessary to capture differ- ent aspects of metaphoric paraphrasing. We re- lease our data and models, as well as our anno- tated results in order to facilitate development of better evaluation metrics.1 1 Introduction Metaphors are ubiquitous in human language, and while humans seem capable of easily understand- ing even complex metaphors, it remains difficult to implement computational methods that capture the depth and breadth of meaning inherent in novel metaphors. While many approaches have been im- plemented for metaphor detection, interpretation, and generation, there remain many open questions about how to incorporate linguistic and cognitive theory into these methods, as well as how to evalu- ate them quickly and effectively. Our task is that of metaphoric paraphrase gen- eration: given a literal input sentence, we aim to generate a semantically similar sentence that is 1Code and data at https://github.com/UKPLab/ conll2021-metaphoric-paraphrase-generation. metaphoric. Specifically, we focus on two possi- ble methods for metaphoric paraphrase generation: free and controlled. In free generation, models are trained using lit- eral and metaphoric pairs, but are not given in- formation regarding which metaphoric meaning is intended. This allows the model the freedom to develop new metaphors that can be creative, novel, and perhaps even inspiring. However, it can lead to noisy generation, which is difficult to evaluate as even randomly generated metaphors can some- times be effective via coercion-like processes. Ad- ditionally, if a generated text is to remain coherent, metaphor generation must be consistent within it, and free generation is thus ill-suited to generating longer coherent metaphoric texts. The other option, controlled generation, in- volves adding additional constraints the genera- tion objective to encourage the model to generate a specific metaphor. This methodology is useful when an intended metaphor is known, or needs to be consistent across a dialogue: the model can be suitably constrained to be consistent. Recent work has explored controlled generation (Stowe et al., 2021); we aim to address if and how adding control …

Content truncated to 500 words