Investigating the Frequency Distortion of Word Embeddings and Its Impact on Bias Metrics
This work addresses a critical artifact in NLP for researchers and practitioners using embedding-based metrics, revealing that frequency distortions can mislead bias assessments, though it is incremental in building on prior frequency encoding findings.
The study investigated how word frequency distorts semantic similarity in static word embeddings like Skip-gram, GloVe, and FastText, finding that high-frequency words show artificially higher similarity, and demonstrated that this distortion can significantly affect gender bias metrics, even reversing bias signs in controlled experiments.
Recent research has shown that static word embeddings can encode word frequency information. However, little has been studied about this phenomenon and its effects on downstream tasks. In the present work, we systematically study the association between frequency and semantic similarity in several static word embeddings. We find that Skip-gram, GloVe and FastText embeddings tend to produce higher semantic similarity between high-frequency words than between other frequency combinations. We show that the association between frequency and similarity also appears when words are randomly shuffled. This proves that the patterns found are not due to real semantic associations present in the texts, but are an artifact produced by the word embeddings. Finally, we provide an example of how word frequency can strongly impact the measurement of gender bias with embedding-based metrics. In particular, we carry out a controlled experiment that shows that biases can even change sign or reverse their order by manipulating word frequencies.