[Question] Clarification on the RIND implementation.

[WWWonderer]

Hello!

Thank you for open sourcing your model! I have some questions regarding the implementation of RIND. From the paper, RIND seems to be the product of [entropy] x [max(attention)] x [boolean(is a stopword)]. However, if I'm not mistaken, your implementation seems to have some additional details in the modifier function of AttnWeightRAG:

# value = attenion * (-log prob)
attns = attentions[tl:tr]
attns = np.array(attns) / sum(attns)
value = [attns[i-tl] * weight[i] * (tr-tl) for i in range(tl, tr)] 
thres = [1 if v > self.hallucination_threshold else 0 for v in value]

Namely you first normalized the attn using attns = np.array(attns) / sum(attns) instead of using raw attention, and you also added the length of the sentence tr-tl into the equation (attns[i-tl] * weight[i] * (tr-tl)). Is this understanding correct? What is the purpose of adding (tr-tl)? Is it used to counter-balance the softmax as longer sentences will tend to have lower values?

There are 2 other points of concern to me:

1. When a word is composed of multiple tokens, you chose to add the attention of each token. Wouldn't this create unfair advantage to multi-token words especially when using the max solver?

2. You chose to average the embedding over multiple heads, wouldn't this defeat the purpose of multi-head attention as they supposedly each represent different criterias?

Thank you very much!