Latent Semantic Clustering of German Verbs
1. Introduction
German verbs are among the most challenging elements for learners because they encode tense, mood, voice, aspect, and a rich system of separable prefixes. Traditional teaching approaches group verbs by morphological criteria (regular vs. irregular, weak vs. strong) or by semantic fields (motion, perception, state). While useful, these categories often ignore the subtle relationships that arise from how verbs are used in real text.
Latent Semantic Clustering (LSC) combines two ideas: latent semantic analysis (LSA) a statistical technique that discovers hidden dimensions in word cooccurrence and clustering an algorithm that groups items based on similarity in that latent space. Applying LSC to a large corpus of German yields clusters that reflect the way native speakers actually associate verbs, revealing patterns that are difficult to capture with handcrafted taxonomies.
2. Methodology
2.1 Corpus Selection
For a robust model we used a balanced mixture of contemporary sources:
- Deutsche Nationalbibliotheks digitised novels ( 200M tokens)
- DW news articles ( 50M tokens)
- Wikipedia German edition ( 70M tokens)
- Opensource forum discussions ( 30M tokens)
The total size is roughly 350million tokens, providing a rich distribution of verb usages across registers.
2.2 Preprocessing
- Tokenisation with
spaCy(German model). - Lemma extraction every verb token is reduced to its infinitive form (e.g., ging
gehen). - Removal of stopwords and punctuation.
- Construction of a verbbycontext matrix where each column represents a sliding window of 5 tokens, capturing immediate syntactic partners.
2.3 Dimensionality Reduction
Singular value decomposition (SVD) was performed on the matrix. The top 300 singular values were retained, balancing noise reduction and semantic richness. This step maps each verb to a 300dimensional vector that approximates its latent meaning.
2.4 Clustering Algorithm
We applied kmeans with the elbow method suggesting k = 25 as a suitable number of clusters. To ensure stability, the algorithm was run 50 times with different seeds and the solution with the lowest withincluster sum of squares was kept.
3. Results Representative Clusters
3.1 Motion & Travel
Verbs such as gehen, laufen, fahren, reisen, wandern and the separable abfahren cluster together. The underlying dimension captures directional or locational change, often cooccurring with prepositions (nach, zu) and adverbs of distance (weit, nah).
3.2 Perception & Cognition
Includes sehen, hren, fhlen, merken, denken, wissen. These verbs appear frequently with mentalstate nouns (Wahrnehmung, Gedanke) and with modal particles that express certainty (doch, eben).
3.3 Communication
Cluster: sagen, erzhlen, mitteilen, fragen, antworten, schreiben, posten. The semantic field is dominated by discourse markers (dass, ob) and indirect speech constructions.
3.4 State & Existence
Verbs: sein, bleiben, liegen, existieren, gehren. These tend to cooccur with adjectives describing properties (gr, alt) and with possessive determiners.
3.5 Change of State
Includes werden, verndern, entwickeln, verbessern, verschlechtern. The cluster reflects processes that move an entity from one state to another, often linked with temporal adverbs (schon, bald).
3.6 Modal & Auxiliary Verbs
Group: knnen, mssen, wollen, sollen, drfen, mgen. Although these verbs have few lexical meanings themselves, they share a strong syntactic pattern: a following infinitive without zu.
3.7 Separable Prefixes Directional
Verbs like aufstehen, einsteigen, ausgehen, weitergehen form a subcluster within motion, emphasizing the role of the prefix in specifying direction or completion.
3.8 Reflexive Actions
Examples: sich erinnern, sich setzen, sich freuen, sich kmmern. The reflexive particle sich creates a distinct distribution because it frequently appears with prepositional phrases indicating personal benefit or internal states.
3.9 Evaluation & Aesthetics
Verbs: schn (as verb schnen), gefllig, leiden, genieen. These are tightly linked with adjectives and adverbials of quality.
3.10 Miscellaneous Small Clusters
Some verbs, because of idiomatic usage, occupy their own tiny clusters (e.g., zaubern, flattern). They still provide useful hints for learners about the contexts where these words appear.
4. Interpretation for Language Learning
Latent semantic clusters can be turned into teaching resources in several ways:
- Vocabulary bundles: Present learners with a set of verbs from the same cluster, accompanied by typical collocates. For the Motion cluster, a learner sees
gehenwith nach Hause,fahrenwith mit dem Zug, etc. - Contextual drills: Use authentic sentences extracted from the corpus to illustrate subtle differences.
laufenoften describes footbased movement, whilefahrenimplies a vehicle. - Contrastive exercises: Pair verbs from adjacent clusters (e.g., perception vs. communication) and ask learners to decide which fits a given context.
- Semantic maps: Visualise the clusters on a 2D plane using tSNE or UMAP, helping learners see the distance between verbs.
Because the clusters are datadriven, they automatically adapt to contemporary usage, covering recent colloquialisms that static schoolbooks may miss.
5. Technical Considerations
Corpus bias. News texts favour formal language, while forums introduce slang. The chosen mix mitigates extreme bias but does not eliminate it; users should be aware of register differences.
Polysemy. Verbs like laufen have multiple senses (run, flow). LSC captures the dominant sense in the corpus; rare senses may be lost unless a sensedisambiguated corpus is built.
Parameter choice. Changing the number of dimensions or clusters can split or merge groups. Researchers are encouraged to experiment with hierarchical clustering to reveal substructures.
6. Future Directions
- Integrate contextual embeddings (e.g., BERTGerman) to capture wordsense dynamics.
- Combine LSC with syntactic parsing to differentiate transitive vs. intransitive uses.
- Develop an interactive web tool where learners can input a verb and receive its cluster, typical collocations, and example sentences.
- Apply the same pipeline to other parts of speech (adjectives, adverbs) for a fully integrated lexical network.
7. Conclusion
Latent semantic clustering offers a statistically grounded view of how German verbs relate to one another in authentic language. The resulting groups align well with intuitive semantic fields while also exposing subtle patterns such as the influence of separable prefixes, reflexivity, and modality. For educators and learners, these clusters provide a practical scaffold for building richer vocabularies, designing meaningful exercises, and keeping pace with the evolving usage of German.
By bridging corpus linguistics and pedagogy, LSC helps turn the complexity of German verb semantics into an accessible, datadriven resource.
Latent Semantic Clustering Of German Verbs and Reference File Download Link
Computational Semantic Lexicon Of French Verbs Of Emotion and Reference File Download Link
Latent Variable Models An Introduction To Factor Path And Structural Equation Analysis and...
German Perfekt Tense For Regular And Irregular Verbs and Reference File Download Link