mask_token for SHAP explainer for text: empty string or [MASK]?

Question

I am using SHAP and, whilst experimenting with different mask tokens, I noticed that they result in quite different results and I can see that from the resulting .base_values of the explanation object. Here I show how I tried out the different mask_tokens:

mask_tokens = ['[MASK]', '', '...']

for mask in mask_tokens:
    masker = shap.maskers.Text(tokenizer=r"\W+", collapse_mask_token=True)
    masker.mask_token = mask
    explainer = shap.Explainer(pipe, masker=masker, seed=1)

    shap_values = explainer(
        some_text
    )

In my case the underlying task is binary text classification, where the data distribution is ~40:60 and the model is a fine-tuned BERT. The base values (== prediction for the mask_token only i.e., pipe(mask_token)) are as follows:

• '' (empty string) leads to [0.41945168, 0.58054835]
• '...' (SHAP default) leads to [0.15069047, 0.8493095 ]
• '[MASK]' (same as the mask token used by the model's tokenizer) leads to [0.28819373, 0.7118063 ]

Is there a right one to use to ensure max level possible of model's explainability? How does one go about deciding which one to use?