Coherence score is a critical arithmetic evaluation metric used to assess the quality and interpretability of topics produced by topic modeling algorithms like LDA, NMF, or BERTopic. A high coherence score suggests that the words within a topic frequently co-occur and are semantically related, indicating a good, high-quality topic.

This score is essential for evaluating different configurations of the model setup, often used in conjunction with perplexity, to determine the optimal hyperparameters (like the number of topics or alpha/beta values). BERTopic specifically aims for strong semantic coherence by capturing contextual meanings.