Abstract
Human Needs categories characterize why an affective event is perceived as positive or negative. For instance, "getting the flu" and "getting fired" are both negative events, but they relate to distinct needs: Health and Finance, respectively. Previous research has focused on assigning events to these categories based on lexical and contextual features. This paper proposes an intermediate semantic classification step to address word sparsity. We develop lightly supervised models to label words with 10 semantic concepts associated with Human Needs and incorporate these labels as features. Our results demonstrate that these semantic features significantly improve both precision and recall, leading to a state-of-the-art F1 score of 65.9%.
Introduction
Understanding affective events—those that impact a person's well-being—is a fundamental task in Natural Language Processing. Beyond identifying sentiment, it is crucial to understand the underlying motivations or "Human Needs" (e.g., Maslow's hierarchy) that an event satisfies or violates. However, event categorization faces a significant "word sparsity" challenge. A model might learn that "fired" relates to Finance but fail to generalize to "downsized" if that specific word was absent from the training data. This work bridges that gap by introducing a semantic layer that groups related words into broader concepts.
Methodology
The proposed approach utilizes a two-stage pipeline. First, a word-level semantic classifier is trained using light supervision. Ten semantic categories were defined: Emotion, Entertainment, Equipment, Food/Drink, Interpersonal, Medical, Mental-Process, Money/Job, People, and Other. Instead of extensive manual labeling, the authors used "seed words" (10–20 per category) to guide the learning process. Three models were evaluated for this task: Logistic Regression, Instance-based learning, and Prototypical Neural Networks (ProtoNets). ProtoNets, a few-shot learning architecture, proved most effective at learning robust class prototypes from limited seeds.
Second, these semantic labels are integrated into a Human Needs categorization model. Events are represented as tuples (Agent, Predicate, Theme, PP), and the semantic classifier provides 50 features (confidence scores and binary indicators) for each tuple element. These features are then used within a co-training framework that iteratively refines the classifier by looking at both the event expression and its surrounding context.
Results and Discussion
The inclusion of semantic features provided a substantial boost to the base event expression classifier. As shown in Table 1, moving from raw word embeddings to a combination of embeddings and semantic features derived from ProtoNets increased the F1 score from 54.8% to 61.9%.
When these features were integrated into the full co-training model, the system achieved an average F1 score of 65.9%, outperforming the previous state-of-the-art (62.4%). A detailed breakdown by category (Table 2) reveals that the semantic features improved performance across every single Human Need category. The most notable improvements were observed in the Health and Cognition categories, where semantic concepts like "Medical" provided strong signals for categorization.
The success of this approach highlights the role of semantics as a regularizer. By forcing the model to map sparse vocabulary to stable conceptual anchors, it becomes more robust to unseen words. For example, recognizing that "epidural" is a Medical concept allows the model to correctly categorize it under Health needs, even without prior exposure to the specific term in an affective context.
Conclusion
This research demonstrates that an intermediate semantic classification step significantly enhances the categorization of affective events into Human Needs. By leveraging few-shot learning techniques like ProtoNets and a co-training framework, the model overcomes word sparsity and achieves superior performance. Future work could address polysemy by utilizing contextualized embeddings like BERT and explore more systematically derived sets of semantic concepts to further refine the categorization process.
Summary updated