Text Embedding Visualization Tool

Overview

This repository demonstrates the application of text embeddings, Uniform Manifold Approximation and Projection (UMAP), and visualization tools such as Plotly Dash to gain insights into a dataset of research paper abstracts and associated metadata.

Installation

Google Colab

To run the demo on Google Colab, follow these steps:

1. Copy the URL of the abstract_summary.ipynb notebook: Notebook URL
2. Navigate to Google Colab.
3. Click on File > Open notebook.
4. Switch to the GitHub tab, paste the copied URL into the search bar, and press Enter.
5. Open the notebook from the search results.

Local Installation

To install and run the demo locally, follow the steps below in the terminal:

$ conda create -n text-embedding-demo python=3.11
$ conda activate text-embedding-demo
$ pip install torch # Use the exact installation command from https://pytorch.org/
$ git clone https://github.com/NoviaIntSysGroup/text-embedding-demo.git
$ cd text-embedding-demo
$ pip install -e .

NOTE: One needs to have conda, python and git installed.

Components

1. 2D UMAP Scatter Plot

• Description: Utilizes UMAP for visualizing data similarity in a 2D scatter plot, where each point corresponds to an individual research paper.
• Interactive Features:
  • Hover: Displays summary/abstract and metadata on hover.
  • Selection: Enables selection of multiple points for comparative analysis.

2. Keyword Bar Graph

• Description: A horizontal bar graph representing the similarity of keywords in the dataset, offering insights into dominant themes.

3. Detailed Data View

• Description: Provides detailed information about a highlighted or selected data entry from the scatter plot.

4. Spider (Radar) Chart

• Description: Visualizes metrics related to the dataset using a radar chart.
• Features:
  • Mean Line: Depicts the average value for each metric across the dataset.
  • Hover Line: Updates to show the metric values of a data point highlighted in the UMAP plot.

Usage

• Explore Clusters: Begin with the 2D UMAP scatter plot to explore data clusters and patterns.
• Inspect Details: Hover over points to view abstracts, metadata, and an updated Spider Chart.
• Group Analysis: Select multiple points for comparative analysis.
• Identify Themes: Abstracts can be color-coded based on their similarity to a specific input query by user or a selected topic sentence. Analyze the Keyword Bar Graph for prevalent themes and leverage the UMAP-generated scatter plot for thematic categorization of abstracts.
• Deep Dive: Use the Detailed Data View and Spider Chart for deeper insights, comparing individual metrics against the dataset mean.

Contributors:

• Johan Westö
• Mikael Manngård
• Ashish Dahal