This repository presents an evaluation and comparison of three popular static word embedding techniques—Singular Value Decomposition (SVD), Continuous Bag of Words (CBOW), and Skip-Gram—on the WordSim353 dataset using Spearman's Rank Correlation. The goal is to assess how well each model captures semantic similarity between words.

• window_size refers to the number of context words around a target word, i.e., ± window_size.

The text corpus is preprocessed before training the embedding models:

• All text is converted to lowercase
• Punctuation is removed
• Stopwords are filtered out

• Word embeddings are saved in model-specific directories.
• Cosine similarity scores computed using the WordSim353-Crowd dataset are saved as: word_similarity.csv

• Pre-trained embeddings can be loaded directly from the appropriate cell.
• Training from scratch is possible but may be computationally expensive.
• Loading embeddings will automatically run evaluation on the WordSim353 dataset.

The effectiveness of SVD, CBOW, and Skip-Gram was checked by analyzing each technique was effective in capturing semantic similarity between word pairs. Evaluation was done using the Spearman's Rank Correlation coefficient between model-computed similarities and human-annotated scores in the WordSim353-Crowd dataset.

• Window Size: 4
• Embedding Dimension: 200

• Spearman Correlation: 0.1985
• p-value: 0.00033
• Words Evaluated: 323 / 353

Observation: Higher embedding dimensions and larger context windows (up to 4) generally yield better correlation scores, suggesting improved semantic representation.

• Window Size: 4
• Embedding Dimension: 200
• Learning Rate: 0.1
• Epochs: 75
• Batch Size: 1024
• Negative Samples: 20

• Spearman Correlation: 0.2212
• p-value: 0.0000607
• Words Evaluated: 323 / 353

Observation: CBOW outperforms SVD and captures moderate levels of semantic similarity.

• Window Size: 4
• Embedding Dimension: 200
• Negative Samples: 20
• Learning Rate: 0.1
• Epochs: 100
• Batch Size: 1000

• Spearman Correlation: 0.2558
• p-value: 0.0000032
• Words Evaluated: 323 / 353

Observation: Skip-Gram outperforms both CBOW and SVD, showing the strongest alignment with human judgments of word similarity.

• Skip-Gram provides the most effective word embeddings in terms of human-perceived similarity.
• CBOW offers moderate performance with faster training time compared to Skip-Gram.
• SVD, while simpler, performs less effectively than neural methods but still captures meaningful relationships.