TL;DR
This paper tackles the challenge of automatically inferring the implicit properties that make similes meaningful, going beyond explicit comparisons like "as cold as Antarctica." They propose a three-step framework that generates candidate properties, evaluates their compatibility with different simile components, and ranks them, achieving significant improvements in inferring these hidden meanings.
What Problem Does This Paper Address?
Similes are a common and powerful linguistic tool, but many are "open similes" where the shared property is not explicitly stated. For example, "my room feels like Antarctica" implies the room is cold, but doesn't say it. Understanding these implicit properties is crucial for deeper natural language understanding, metaphor interpretation, and even sentiment analysis, as the implied property can reveal positive or negative connotations. The authors aim to automatically infer these implicit properties, a task that has been largely unexplored computationally.
Methods
The paper introduces a three-step framework for inferring implicit properties:
- Candidate Property Generation: They generate potential properties by looking at the "vehicle" (e.g., "Antarctica") and "event" (e.g., "feels") of the simile. This involves several methods:
- Harvesting adjectives and adverbs directly modifying the vehicle or event.
- Extracting explicit properties mentioned in similar comparative structures (e.g., "cold like Antarctica").
- Leveraging dictionary definitions of the vehicle and event terms.
- Using Pointwise Mutual Information (PMI) to find words that co-occur frequently with the vehicle or event.
- Employing word embeddings (trained on their Twitter corpus) to find words semantically similar to the vehicle or event.
- Candidate Property Evaluation: Each generated candidate property is then evaluated for its compatibility with the other component of the simile (e.g., a property generated from the vehicle is checked against the event). This is done using PMI and cosine similarity of word embedding vectors.
- Aggregated Ranking: Finally, the candidates from all generation methods are combined and ranked. They use a harmonic mean of the ranks from individual methods, rewarding properties that consistently rank high across different approaches. This is compared to a simple voting method.
Datasets
The authors created a new dataset of open similes from approximately 140 million English tweets collected between February 2013 and April 2014. This dataset was filtered to focus on figurative similes with implicit properties. A gold standard set of properties was developed using seven human annotators on Mechanical Turk, who provided up to two properties per simile. This dataset was split into a development set (29%) and a test set (71%).
Key Insights and Ideas
- Vehicles are richer sources: The study found that the "vehicle" component of a simile (e.g., "Antarctica") is a much more effective source for generating candidate properties than the "event" component (e.g., "feels").
- Interpretive diversity matters: Similes with a wider range of plausible interpretations (higher "interpretive diversity" as measured by Shannon's entropy) are significantly harder to infer properties for. This suggests that ambiguity in interpretation directly correlates with inference difficulty.
- Combining sources is key: No single candidate generation method is sufficient. Combining properties from multiple sources (syntactic patterns, dictionaries, embeddings) leads to much higher coverage of similes.
- Contextual evaluation improves ranking: Evaluating candidate properties not just in isolation but also for their compatibility with the other simile component (event/vehicle) significantly improves the ranking of plausible properties.
Key Results
- The proposed three-step framework achieved an MRR (Mean Reciprocal Rank) of 0.33 for the gold standard (Gd) and 0.41 for the expanded gold standard (Gd+WN), outperforming individual methods and a simple voting approach.
- The aggregated ranking scheme successfully ranked an acceptable property in the Top 1 position for 27% of similes (using Gd+WN) and within the Top 5 positions for 58% of similes.
- The system performed best on similes with low interpretive diversity, achieving an MRR of 0.50 compared to 0.31 for similes with high interpretive diversity.
- The combined set of properties from all candidate generation methods (Top 30) achieved 88% coverage of the similes.
Limitations and Future Work
- Limited event contribution: The "event" component of similes, especially perception verbs like "feels" or "is," often provides weak signals for property generation.
- Multi-word vehicles: The current approach primarily focuses on single-word vehicles and leaves multi-word vehicle phrases (e.g., "stepping into a hurricane") for future work.
- Future work: The authors plan to use the inferred properties to improve affective polarity recognition in similes. They also suggest exploring how to better leverage the "event" component and handle more complex simile structures.
Summary updated