We posit that low structural correspondence between utterances and programs hurts generalization.
To improve correspondence, we study reversible and lossy intermediate representations and identify key principles for designing effective representations for new formalisms.
2/4
What is the impact of intermediate representations on compositional generalization?
We find them to be surprisingly effective in improving semantic parsing generalization for pre-trained LMs!
w/ @ptshaw2 @mchang21 @kelvin_guu @IcePasupat Yuan Zhang
arxiv.org/pdf/2104.07478.pdf
1/4
4/ SpanBasedSP performs similarly to strong seq2seq baselines on random splits, but dramatically improves performance compared to baselines on splits that require compositional generalization.
From 69.8 → 95.3 average accuracy on SCAN, CLOSURE and GeoQuery.
3/ To be comparable with seq2seq models, we train from programs, without access to gold trees, treating trees
as latent variables. We also parse a common class of non-projective trees through an extension to standard CKY.
1/ Does a span-based parser perform better than seq2seq models in terms of compositional generalization in semantic parsing?
Seems like it!
Check out our new paper: arxiv.org/pdf/2009.06040.pdf, Joint work with @JonathanBerant
