StabilityAnalysisRandomNoise

Evaluate robustness of embeddings models to random noise introduced by using a probability parameter to choose random locations in the text to apply random perturbations. These perturbations include:

• Swapping two adjacent words
• Introducing a random typo in a word
• Deleting a random word
• Inserting a random word at a random position

Purpose: The purpose of the stability analysis is to evaluate the robustness of a text embeddings model to random noise. Random perturbations such as swapping adjacent words, introducing random typos, deleting random words, or inserting random words at random positions in the text are introduced, to gauge the model’s performance and stability.

Test Mechanism: The test mechanism includes a series of function-defined random perturbations like swapping two words random_swap, introducing a typo in a word introduce_typo, deleting a random word random_deletion, and inserting a random word at a random position random_insertion. A probability parameter determines the likelihood or frequency of these perturbations within the text data.

The perturb_data function initally tokenizes the string data based on spaces then applies selected random perturbations based on the provided probability for each word in the text.

Signs of High Risk: - High error rates in model predictions or classifications after the introduction of the random noise - Greater sensitivity to certain types and degrees of noise such as typographical errors, insertion or deletion of words - Significant change in loss function or accuracy metric - Inconsistency in model outputs for slightly perturbed inputs

Strengths: - Measures model robustness against noise thereby reflecting real-world scenarios where data may contain errors or inconsistencies. - Easy to implement with adjustable perturbation severity through probability parameter. - Enables identification of model sensitivity to certain noise types, providing grounding for model improvement. - Useful for testing models designed to handle text data.

Limitations: - The test might not be effective for models that have been designed with a high resistance to noise or models that are inherently designed to handle such perturbations. - Pseudo-randomness may not accurately represent real-world distribution of noise or typographical errors, which could be biased towards certain types of errors or malware injections. - Highly dependent on the probability parameter to introduce noise, with artificial adjusting required to achieve an optimal balance. - Only validates the model’s performance against noise in input data, not its ability to capture complex language structures or semantics. - Does not guarantee the model’s performance in new, unseen, real-world data beyond what is represented by the noise-introduced test data.