TLDR: A new semantic-driven topic modeling framework uses transformer embeddings (Sentence-BERT), dimensionality reduction (UMAP), clustering (HDBSCAN), and topic extraction to analyze virtual brainstorming sessions. It effectively identifies coherent themes, filters noise, and provides interpretable insights into group creativity, outperforming traditional methods in topic coherence.
Virtual brainstorming has become a cornerstone of modern collaboration, yet the sheer volume and varied nature of ideas generated often make it challenging to extract meaningful insights. Traditional manual analysis is time-consuming and subjective, highlighting a clear need for automated tools to help evaluate group creativity effectively.
Addressing this challenge, a new research paper introduces a semantic-driven topic modeling framework designed to analyze creativity in virtual brainstorming sessions. This innovative approach integrates several advanced computational techniques to uncover coherent themes from brainstorming transcripts, while also filtering out irrelevant information and identifying unique, outlier ideas.
The Core Framework
The proposed framework is built upon four interconnected components:
Transformer-based Embeddings (Sentence-BERT): This initial step converts each idea or sentence into a dense numerical vector. Sentence-BERT is crucial because it captures the contextual meaning of words and sentences, ensuring that ideas with similar meanings are represented closely in a digital space. This goes beyond simple word matching to understand the true semantic relationships between ideas.
Dimensionality Reduction (UMAP): The high-dimensional vectors produced by Sentence-BERT can be complex for subsequent analysis. UMAP (Uniform Manifold Approximation and Projection) is used to reduce these vectors into a lower-dimensional space. This process makes computations more efficient and helps to better separate distinct groups of ideas, all while preserving the underlying semantic structure.
Clustering (HDBSCAN): Once the dimensions are reduced, HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) groups these ideas into semantically coherent clusters. A key advantage of HDBSCAN is its ability to identify clusters of varying shapes and densities, and importantly, it labels low-density points as outliers. This means the framework can effectively filter out noise and irrelevant ideas, focusing on the most meaningful contributions.
Topic Extraction with Refinement: For each identified cluster, the framework then extracts the most representative words to define its underlying topic. This involves creating a vocabulary for each cluster, generating word embeddings, and computing the average semantic similarity of each word within its cluster. The top ‘k’ words with the highest similarity are selected as topic words. A refinement step further merges similar topics to reduce redundancy and achieve a user-specified number of topics.
Why This Approach Matters
This semantic-driven framework offers significant advantages over older methods like Latent Dirichlet Allocation (LDA) and Embedding Topic Models (ETM). The researchers evaluated their model on structured Zoom brainstorming sessions where student groups worked on improving their university. The results showed that their model achieved a higher topic coherence score of 0.687 (CV) on average, significantly outperforming established baselines. This higher coherence means the extracted topics are more meaningful and easier for humans to interpret.
Beyond just performance, the model provides valuable insights into both convergent and divergent thinking patterns within group creativity. Convergent thinking involves focusing on a single solution, while divergent thinking explores many possibilities. By identifying distinct topics and their relationships, the framework helps understand how groups explore ideas in depth and breadth.
For instance, the study identified topics such as “education,” “bookstore,” “scholarship,” “parking,” and “dining,” among others, from the student brainstorming sessions. The framework can even highlight relationships between topics, like how “schedule” might relate to “class” or “dining.”
Also Read:
- Scaling Qualitative Research: The HICode Approach to Deep Text Analysis
- CLIMB: A New Framework for Data-Driven Occupation Taxonomies
Looking Ahead
The researchers plan to enhance the framework further by integrating Explainable AI (XAI) techniques. This will provide even greater transparency, showing which words and sentences most strongly influence topic assignments and coherence. Future work may also explore incorporating multimodal signals, such as audio and video cues from virtual meetings, and applying the framework to other domains.
This work represents a significant step forward in automating the analysis of collaborative ideation, offering a scalable and interpretable tool for studying creativity in synchronous virtual meetings. You can read the full research paper for more details here: Semantic-Driven Topic Modeling for Analyzing Creativity in Virtual Brainstorming.
