Closed pengzhansun closed 2 years ago
Hi,
The code is general, you just need to call it with different parameters, for the different experiments. If you look at scripts/test_lm.sh, for example, you will see the following comment:
Modify the test_masking_policy to "seen_verbs" (if testing seen verbs),
"seen_nouns" (if testing seen nouns), "new_combo_seen_noun_seen_verb_merge"
(if testing new compositions) or "seen_combo_seen_noun_seen_verb_merge"
(if testing seen compositions).
The parameter test_masking_policy
ends up affecting the way the masks are created (this is, which words are evaluated), in the function compute_item_mask.
An interesting paper. I'm exactly interested in the compositionality mentioned in your paper, but I cannot find codes or related data split in composition. Is it available to share this part?