[Paper Review] Generating similes effortlessly like a Pro: A Style Transfer Approach for Simile Generation

Tuhin Chakrabarty, Smaranda Muresan+1 more|arXiv (Cornell University)|Sep 18, 2020

Natural Language Processing Techniques34 references4 citations

TL;DR

This paper proposes SCOPE, a style transfer approach for generating creative similes by fine-tuning BART on a large automatically constructed parallel corpus of literal sentences and their simile counterparts. Using common sense knowledge from COMET, the method generates novel similes that are rated higher than those by literary experts 37% of the time and outperforms baselines 68-82% of the time, demonstrating strong performance in figurative language generation.

ABSTRACT

Literary tropes, from poetry to stories, are at the crux of human imagination and communication. Figurative language such as a simile go beyond plain expressions to give readers new insights and inspirations. In this paper, we tackle the problem of simile generation. Generating a simile requires proper understanding for effective mapping of properties between two concepts. To this end, we first propose a method to automatically construct a parallel corpus by transforming a large number of similes collected from Reddit to their literal counterpart using structured common sense knowledge. We then propose to fine-tune a pretrained sequence to sequence model, BART~\cite{lewis2019bart}, on the literal-simile pairs to gain generalizability, so that we can generate novel similes given a literal sentence. Experiments show that our approach generates $88\%$ novel similes that do not share properties with the training data. Human evaluation on an independent set of literal statements shows that our model generates similes better than two literary experts extit{37\%}\footnote{We average 32.6\% and 41.3\% for 2 humans.} of the times, and three baseline systems including a recent metaphor generation model extit{71\%}\footnote{We average 82\% ,63\% and 68\% for three baselines.} of the times when compared pairwise.\footnote{The simile in the title is generated by our best model. Input: Generating similes effortlessly, output: Generating similes extit{like a Pro}.} We also show how replacing literal sentences with similes from our best model in machine generated stories improves evocativeness and leads to better acceptance by human judges.

Motivation & Objective

• To address the lack of training data for simile generation by automatically constructing a large parallel corpus of literal sentences and their simile equivalents.
• To frame simile generation as a style transfer task, transforming literal descriptions into figurative comparisons using shared properties.
• To improve the quality and creativity of generated similes by incorporating structured common sense knowledge into the training process.
• To evaluate the impact of generated similes on downstream creative text generation, such as story evocativeness and human preference.

Proposed method

• Automatically collect similes from Reddit using the keyword 'like a' to form an initial set of parallel pairs.
• Transform each simile into its literal counterpart by removing the comparator 'like' and replacing the vehicle with its associated property using COMET, a commonsense knowledge model.
• Fine-tune the BART sequence-to-sequence model on the resulting [literal, simile] pairs to learn the style transfer from literal to figurative expression.
• Leverage structured common sense knowledge from COMET to ensure that the generated similes maintain semantic coherence and meaningful property mappings.
• Apply the trained model to post-process machine-generated stories, replacing literal descriptions with generated similes to enhance evocativeness.
• Use human evaluation on Amazon Mechanical Turk to compare model outputs against literary experts and baseline systems.

Experimental results

Research questions

• RQ1 Can a large-scale parallel corpus of literal sentences and their simile equivalents be automatically constructed using distant supervision and common sense knowledge?
• RQ2 Can fine-tuning a pre-trained sequence-to-sequence model like BART on such a corpus generate high-quality, novel similes that are not simply memorized from training data?
• RQ3 How do model-generated similes compare to those written by human literary experts in terms of quality and creativity?
• RQ4 To what extent do generated similes improve the evocativeness and human acceptance of machine-generated stories?

Key findings

• The proposed method generated 88% of similes that did not share any properties with training data, indicating strong generalization and novelty.
• Human evaluation showed that the model's similes were preferred over those of two literary experts 37% of the time on average.
• The model outperformed three baseline systems, including a recent metaphor generation model, in pairwise comparisons 68% to 82% of the time.
• Post-processing machine-generated stories with model-generated similes significantly improved their evocativeness and was preferred by human judges.
• The integration of common sense knowledge via COMET was critical for generating semantically meaningful and contextually relevant similes.
• The approach demonstrates that style transfer with commonsense-augmented data can effectively generate creative figurative language, opening new directions for NLG in creative writing.