Semantic Antipode Project Walkthrough

I implemented the “Semantic Antipode” finder, which calculates the mathematical opposite of words in the GloVe vector space relative to the global mean.

Setup & Execution

1. Environment: A virtual environment was created, and gensim, numpy, scipy were installed.
2. Execution: Run the script from the terminal:

   source venv/bin/activate
   python3 antipode.py

Implementation Details

• Memory Constraint: The script uses gensim to download the glove-wiki-gigaword-300 model but effectively loads only the top 100,000 vectors handling my 8GB RAM limit.
• Math:
  • Global Mean ($\mu$): Calculated as the centroid of all 100,000 vectors.
  • Antipode ($V'$): Calculated as $V' = 2\mu - V$ (Reflection across $\mu$).

Verification Results

I tested the script with the following words:

Input Word	Top “Antipode” Result (Similarity)	Notes
happy	keyrates (0.477)	The mathematical antipode lands in a sparse/noisy region.
king	afp02 (0.437)	Similar to happy, the reflection points to technical tokens or noise.

[!NOTE] Interpretation: The “true opposite” defined by strictly reflecting across the global mean often points to the “empty” parts of the high-dimensional space or clusters of rare/technical tokens, rather than linguistic antonyms (which are usually similar in vector space). This confirms the script is performing the requested mathematical operation, even if the linguistic result is abstract.