Abstract
Comparisons are common linguistic devices used to indicate the likeness of two things. Often, this likeness is not meant in the literal sense—for example, “I slept like a log” does not imply that logs actually sleep. In this paper we propose a computational study of figurative comparisons, or similes. Our starting point is a new large dataset of comparisons extracted from product reviews and annotated for figurativeness. We use this dataset to characterize figurative language in naturally occurring comparisons and reveal linguistic patterns indicative of this phenomenon. We operationalize these insights and apply them to a new task with high relevance to text understanding: distinguishing between figurative and literal comparisons. Finally, we apply this framework to explore the social context in which figurative language is produced, showing that similes are more likely to accompany opinions showing extreme sentiment, and that they are uncommon in reviews deemed helpful.
Introduction
While Natural Language Processing (NLP) has made significant strides in detecting metaphors (e.g., "life is a journey"), similes—comparisons using "like," "as," or "than"—have been largely overlooked. The challenge lies in the lack of a simple syntactic rule to separate literal comparisons from figurative ones. For instance, "Sterling is cheaper than gold" is a literal statement about price, whereas "Her voice shines brighter than gold" is a figurative statement about timbre. Distinguishing between these is crucial for text understanding, sentiment analysis, and building common-sense knowledge bases. This paper bridges this gap by characterizing the linguistic patterns of similes and exploring the social context in which they appear.
Dataset
The authors introduced a new dataset of comparisons extracted from Amazon Product Reviews (spanning categories like Books, Music, Electronics, and Jewelry) and the WaCky/WaCkypedia corpora to provide broader linguistic variety. The dataset was annotated via Amazon Mechanical Turk. Out of 2,400 candidates, the researchers curated a high-confidence set of 273 figurative and 609 literal comparisons where all human annotators agreed. This dataset is made publicly available for the research community.
Linguistic Characterization
The researchers identified several key linguistic patterns that distinguish similes from literal comparisons by analyzing the Topic (the thing being described), the Vehicle (the thing it's compared to), and the shared Property:
- The "In-Domain" Rule: Literal comparisons tend to stay within their domain. If reviewing a camera, a literal comparison will likely use another electronic device as the vehicle. Figurative comparisons almost always reach outside the domain for a vehicle that is less specific to the context but highly "imageable."
- Semantic Distance: In literal comparisons, the topic and vehicle are usually semantically similar (e.g., two metals). In similes, they are often far apart (e.g., a voice and a metal).
- Imageability: Figurative comparisons tend to use properties that are more imageable—words that easily evoke a mental image.
Classification and Results
Using these insights, the authors developed a logistic regression model to binarize comparisons into "literal" or "figurative." The model incorporated domain-specific cues (specificity), semantic similarity using distributional semantics, psycholinguistic features (imageability and concreteness), and syntactic cues (such as the use of indefinite articles).
The "Full" model achieved 88% accuracy (AUC of 0.94), which is remarkably close to the estimated human accuracy of 96%. Notably, the linguistically-inspired features alone outperformed a massive "Bag of Words" baseline, proving that the underlying theory of how similes work is computationally sound. Furthermore, tests showed that state-of-the-art metaphor detection systems performed poorly on this task, confirming that similes require a specialized approach.
Social Context and Conclusion
The study reveals a striking social insight: similes are highly correlated with extreme sentiment (1-star or 5-star reviews). However, as a review's "helpfulness" rating increases, the density of figurative language decreases. This suggests that users prefer literal, factual comparisons when looking for buying advice.
While the study focused on noun-noun comparisons with adjectival properties, future work could explore comparisons involving proper nouns or adverbs. Additionally, investigating the causal effects of figurative language—whether similes inherently reduce helpfulness or are simply a byproduct of emotional feedback—remains an open question for the field.
Summary updated