Intertopic Distance Map

The map uses Principal Coordinate Analysis (PCoA) to project topic-topic distance matrices into two dimensions. The distance between any two topic centroids is calculated using Jensen-Shannon divergence between their word distributions. This preserves the original high-dimensional inter-topic relationships as faithfully as possible in a 2D plane, allowing you to visually identify which topics are semantically proximal and which are distinct.

Each topic is represented as a circle whose area is proportional to the marginal topic prevalence in the corpus. Larger circles indicate topics that appear more frequently across documents. This visual encoding provides an immediate, intuitive sense of which themes dominate the corpus without needing to inspect numerical tables. A large, isolated circle suggests a dominant, well-defined theme; a small circle overlapping others suggests a niche or less coherent topic.

The spatial proximity of circles directly reflects semantic similarity. Overlapping circles indicate topics that share many high-probability words and may be redundant. Isolated circles represent highly distinct, interpretable topics. This visual diagnostic helps determine if the chosen Number of Topics (K) is appropriate: excessive overlap suggests K is too high, while a few massive circles suggest K is too low.

Clicking a topic circle reveals a horizontal bar chart of its salient terms, ranked by a tunable relevance metric (λ). This metric balances a term's frequency within the topic against its exclusivity:

• λ = 1: Ranks by purely by probability within the topic (frequent, potentially generic words)
• λ = 0: Ranks by lift over the corpus-wide background frequency (highly exclusive, potentially rare words) Adjusting λ helps analysts interpret what makes a topic unique.

The map is fully interactive, enabling a systematic diagnostic workflow:

• Hover over circles to see the top terms for each topic instantly
• Click a circle to lock the term bar chart and inspect word distributions in detail
• Identify overlapping clusters that may indicate the need to merge topics or reduce K
• Validate that isolated topics correspond to semantically coherent themes by examining their top-ranked exclusive terms at λ ≈ 0.3

While the Intertopic Distance Map provides a qualitative visual assessment, it complements quantitative metrics like Topic Coherence and C_V Coherence. A well-structured map with distinct, non-overlapping circles typically correlates with high coherence scores. Conversely, a map showing a diffuse cloud of overlapping small circles often indicates low semantic interpretability, prompting a re-evaluation of preprocessing steps, the number of topics, or the choice of model hyperparameters like Alpha and Beta.