---

name: representations-labeling description: "Fine-tune BERTopic topic representations, labels, and multi-aspect outputs." disable-model-invocation: true metadata: disco-role: operating package: "BERTopic" package-version: "0.17.4" parent-skill: "bertopic" license: MIT

representations-labeling

Use this sub-skill when the task is about post-hoc topic representations, topic labeling, multi-aspect outputs, or prompt-based label generation.

Route here for

• KeyBERTInspired, MaximalMarginalRelevance, and other keyword rerankers.
• PartOfSpeech keyword refinement with spaCy patterns.
• ZeroShotClassification against a candidate label list.
• Prompt-based labelers: TextGeneration, OpenAI, LiteLLM, LangChain, Cohere, and LlamaCPP.
• VisualRepresentation for image-centered or multimodal topic aspects.
• Custom BaseRepresentation classes, including chained representation lists and multi-aspect dicts.
• generate_topic_labels, set_topic_labels, topic_aspects_, get_topic(full=True), and get_topics(full=True) when the question is about labels or aspect views.

Route elsewhere

• Embedding backend selection, model download, or encoder troubleshooting: use the embeddings-backends sub-skill.
• c-TF-IDF, CountVectorizer, online vocabulary updates, or tokenization choices: use the vectorizers-ctfidf sub-skill.
• Core fit/transform/partial_fit, clustering, topic mutation, or topic reduction: use the topic-modeling sub-skill.
• Plotting, dashboarding, and figure interpretation belong in the analysis/visualization sub-skill.
• Save/load, hub publishing, and serialization format choices belong in the serialization sub-skill.

Operating references

1. references/api-reference.md
2. references/workflows.md
3. references/troubleshooting.md
4. scripts/smoke_representations.py

Minimal decision flow

• If you only want to soften or diversify the default keywords, start with KeyBERTInspired or MaximalMarginalRelevance.
• If you need noun-phrase or POS filtering, use PartOfSpeech and keep a spaCy model available.
• If you already know the labels, use ZeroShotClassification before falling back to generated labels.
• If you want generated labels, use TextGeneration, OpenAI, LiteLLM, LangChain, Cohere, or LlamaCPP with a prompt that names the label format clearly.
• If you need several views of the same topic, use a dict representation_model with a Main pipeline plus aspect names, then inspect topic_aspects_.
• If you need a chained label flow, pass a list of representation models in order; each tuner sees the previous output.
• If you need KeyBERTInspired to use the top-level embeddings shortcut, keep it at the top level instead of hiding it inside a list or aspect dict.
• If you only need to rename topics after fitting, prefer generate_topic_labels or set_topic_labels rather than retraining.

Verification anchors

• A chained representation model that reranks keywords and then emits a concise topic label.
• Label-setting edge cases for outlier and non-outlier topic ids.
• Multi-aspect topic retrieval through topic_aspects_ and aspect-aware generate_topic_labels(aspect=...).