ink-usc / hierarchical-explanation-neural-sequence-models Goto Github PK

Source code for "Towards Hierarchical Importance Attribution: Explaining Compositional Semantics for Neural Sequence Models", ICLR 2020.

License: MIT License

Python 99.93% Shell 0.07%

hierarchical-explanation-neural-sequence-models's People

Contributors

Stargazers

Watchers

Forkers

mhsamavatian linktopast1990 es94129 agiats hunterjiang97 codefly13 jstremme danield21

hierarchical-explanation-neural-sequence-models's Issues

Using CD 2019, would we still see the computation of β involves the γ term of a specific input sentence?

        # cd 2019
        rel_hidden = activation(rel_input)  - activation(rel_input - rel_input) # rel_hidden = activation(rel_input) ???

Referring to ACD paper Eq. 10, it seems now the computation of β does not involve the γ term, right?

The SST-2 dataset for explanation is different from the SST-2 for training.

Should we perform softmax before computing the prediction difference?

logits_enb = self.model(inp_enb, segment_ids[:, :inp_enb.size(1)], inp_mask_enb)
logits_ex = self.model(inp_ex, segment_ids[:, :inp_ex.size(1)], inp_mask_ex)
contrib_logits = logits_enb - logits_ex  # [1 * C]

Hi, the logits_enb and logits_ex may have very different values, while softmax(logits_enb) and softmax(logits_ex) remains similar.
Should we perform softmax before computing the prediction difference?

Recommend Projects