TL;DR
Researchers have developed a weakly supervised framework to automatically induce a massive lexicon of over 175,000 "affective events"—everyday occurrences that implicitly trigger positive or negative emotions (like "getting a job" vs. "breaking an arm"). By representing events as rich structures and optimizing for semantic consistency across a massive graph, their model significantly outperforms existing sentiment lexicons and traditional learning methods.
Introduction
Most sentiment analysis tools are proficient at identifying explicit emotions (e.g., "I am happy") or opinionated words (e.g., "This is a terrible movie"). However, they struggle with affective events: situations that carry implicit emotional weight based on cultural norms and personal experience. For instance, "I went to the beach" is generally positive, while "I went to prison" is negative. To a standard sentiment analyzer, both look like a neutral "went to [location]" structure. The challenge is that the polarity of an event depends on the complex interaction between a predicate and its participants. This research aims to bridge this gap by automatically learning which events are stereotypically positive or negative from a massive, unannotated corpus of personal stories.
Methodology
The authors propose an iterative learning framework based on a Semantic Consistency Model.
Event Extraction
Events are represented as a 4-tuple: <Agent, Predicate, Theme, Prepositional Phrase>. This structure allows the model to distinguish between nuanced situations where the prepositional phrase or theme flips the polarity.
Graph Construction and Optimization
The core of the framework is a graph where nodes represent events and their individual components (words). These nodes are linked using three types of edges:
- Semantic Similarity: Connecting events with similar meanings (e.g., "have party" and "have celebration") using GloVe embeddings.
- Semantic Opposition: Connecting negated events to their affirmative counterparts (e.g., "I won" vs. "I did not win").
- Shared Components: Connecting an event to its constituent parts (e.g., the event "dropped phone" is linked to the component "phone").
After a noisy initialization using existing sentiment tools (like the MPQA lexicon), the model employs an iterative algorithm to minimize KL-divergence across the graph. This ensures that similar events have similar scores, opposing events have opposite scores, and events align with the polarities of their components.
Key Results
- Scale: The model successfully induced a collection of over 175,000 affective events.
- Accuracy: The full model achieved a macro-averaged F1 score of 71.4%, a significant jump from the 66.3% achieved by the best baseline (a combination of existing lexicons and classifiers).
- Precision vs. Recall: By adjusting the polarity threshold, the model can produce high-precision lexicons. At a threshold of 0.6, the model identifies over 69,000 events with over 93% precision for both positive and negative categories.
- Ablation Study: Results showed that adding "component" nodes (linking events to their individual words) provided the most significant boost in recall, helping the model generalize to previously unseen events.
Discussion and Future Work
The finding that nearly 40% of events in personal narratives carry affective weight highlights the importance of this work for tasks like narrative understanding and sarcasm detection. While the model achieves high precision, future work will focus on improving recall for events without strong semantic links to initial seeds and refining the handling of complex linguistic structures like negation and contextual nuance.
Summary updated