[Paper Review] MOVER: Mask, Over-generate and Rank for Hyperbole Generation

Yunxiang Zhang, Xiaojun Wan|arXiv (Cornell University)|Sep 16, 2021

Multimodal Machine Learning Applications4 citations

TL;DR

This paper proposes MOVER, an unsupervised method for generating hyperbolic paraphrases from literal sentences using a novel 'mask, over-generate, and rank' pipeline. It introduces HYPO-XL, the first large-scale English hyperbole corpus with 17,862 sentences, and demonstrates through automatic and human evaluation that MOVER effectively generates hyperbolic text while preserving meaning, outperforming baseline models.

ABSTRACT

Despite being a common figure of speech, hyperbole is under-researched in Figurative Language Processing. In this paper, we tackle the challenging task of hyperbole generation to transfer a literal sentence into its hyperbolic paraphrase. To address the lack of available hyperbolic sentences, we construct HYPO-XL, the first large-scale English hyperbole corpus containing 17,862 hyperbolic sentences in a non-trivial way. Based on our corpus, we propose an unsupervised method for hyperbole generation that does not require parallel literal-hyperbole pairs. During training, we fine-tune BART to infill masked hyperbolic spans of sentences from HYPO-XL. During inference, we mask part of an input literal sentence and over-generate multiple possible hyperbolic versions. Then a BERT-based ranker selects the best candidate by hyperbolicity and paraphrase quality. Automatic and human evaluation results show that our model is effective at generating hyperbolic paraphrase sentences and outperforms several baseline systems.

Motivation & Objective

• To address the scarcity of large-scale hyperbolic text data for training NLP models.
• To develop an unsupervised method for generating hyperbolic paraphrases from literal sentences without requiring parallel literal-hyperbole pairs.
• To balance content preservation and hyperbolic exaggeration in generated text.
• To enable controllable, interpretable, and flexible hyperbole generation using a three-stage pipeline.

Proposed method

• Construct HYPO-XL, a large-scale, non-parallel corpus of 17,862 hyperbolic sentences collected via weakly supervised retrieval and human-annotated refinement.
• Fine-tune BART to perform span infilling on masked hyperbolic spans within sentences from HYPO-XL, promoting hyperbolic generation during training.
• During inference, mask parts of an input literal sentence and over-generate multiple hyperbolic variants using the fine-tuned BART model.
• Use a BERT-based ranker to select the best candidate based on both hyperbolicity and paraphrase quality.
• Frame hyperbole generation as a sentence editing task rather than a style transfer or paraphrase generation task, enabling better control over exaggeration.
• Leverage the 'overgenerate-and-rank' paradigm to improve output quality without requiring parallel training data.

Experimental results

Research questions

• RQ1 Can a large-scale, non-parallel hyperbole corpus be constructed effectively using weak supervision and human-in-the-loop refinement?
• RQ2 Can a masked sequence-to-sequence model be fine-tuned to generate hyperbolic text without parallel literal-hyperbole pairs?
• RQ3 Does the overgenerate-and-rank pipeline outperform end-to-end or baseline methods in generating meaningful, hyperbolic paraphrases?
• RQ4 Can a BERT-based ranker effectively balance hyperbolicity and paraphrase quality in generated outputs?

Key findings

• The proposed HYPO-XL corpus contains 17,862 high-quality hyperbolic sentences, making it the largest such corpus in English to date.
• The BERT-based hyperbole detection model fine-tuned on HYPO achieved 80% accuracy on the test set, demonstrating strong performance for hyperbole detection.
• MOVER outperforms multiple baseline models in both automatic and human evaluations, showing superior performance in generating hyperbolic paraphrases.
• Human evaluation confirms that MOVER-generated sentences achieve a favorable trade-off between hyperbolicity and content preservation.
• The overgenerate-and-rank strategy enables better interpretability and flexibility compared to end-to-end generation models.