Introduction
In the German language, "particle verbs" are formed by combining a base verb (BV) with a prefix (PV). For instance, strahlen ("to beam") becomes anstrahlen ("to smile at"). While some of these semantic shifts are predictable and "regular," many are irregular or opaque. This research explores a computational model designed to categorize these shifts by looking beyond standard word definitions and incorporating "affective" features—the emotional tone and abstractness of a verb's context.
The Analogy Framework
The core of this study treats the semantic shift as an analogy task: $BV_1 : PV_1 :: BV_2 : PV_2$. If two similar base verbs undergo the same prefixation and result in similar meanings, the shift is considered regular. This framework allows researchers to quantify how much a verb's meaning deviates from its base when a prefix is added.
Methodology
The researchers compared several modeling approaches to solve this analogy task:
- Distributional Similarity: Standard word vectors trained on a 12-billion-token German corpus to measure cosine similarity.
- Generalization Models: Grouping verbs into "neighborhoods" using GermaNet or unsupervised k-Means clustering to see if local semantic context improves prediction.
- Affective Context Models: This innovative approach measures the "affective" flavor of the 500 most associated context words for each verb across nine dimensions, including emotions (anger, fear, joy), sentiment (valence, arousal), and concreteness.
Data and Categorization
A new gold-standard dataset of 685 analogy questions was created, focusing on four frequent prefixes: ab, an, auf, and aus. These analogies were categorized into four distinct types based on the nature of the shift:
- COMP: Both pairs are compositional (no shift).
- ASYMCOMP: Only one pair shifts.
- SHIFTDIFF: Both shift, but the base verbs are semantically unrelated.
- SHIFTREG: The "classic" regular analogy where both shift and the base verbs are similar.
Key Findings
The study revealed that "regular" shifts (SHIFTREG), which are the primary focus of most linguistic studies, are actually the rarest category, representing only 8.5% of the dataset. As shown in the data distribution, most shifts are idiosyncratic or asymmetrical.
Crucially, the model using only contextual affect outperformed all others, achieving 58.39% accuracy in the 4-class task and 78.54% in a binary shift-detection task. This suggests that the emotional "flavor" of surrounding words is a more potent indicator of a meaning shift than the verb's inherent definition. For example, as a verb shifts from a concrete action to a metaphorical one, its surrounding vocabulary typically becomes more abstract and emotionally charged. Ablation studies showed that the emotion "fear" was a particularly strong signal for non-literal shifts.
Conclusion
By integrating affective context into semantic similarity models, researchers can better track the complex, often irregular shifts in verb meaning. This approach demonstrates that semantic change is deeply tied to the emotional and abstract context of usage, holding promise for broader applications in metaphor detection and the study of historical language evolution.
Summary updated