Two papers from me and colleagues have been accepted by ACL.

Unsupervised Learning of PCFGs with Normalizing Flow looks at how context embeddings like ELMo can be used with unsupervised PCFG induction models to incorporate morphological, semantic and contextual information.

Variance of average surprisal: a better predictor for quality of grammar from unsupervised PCFG induction tackles the problem of objective function in unsupervised PCFG induction. Data likelihood has been used for induction, but the correlation between it and performance is weak. What other indicator may be better at predicting grammar induction performance?

Very excited to present and discuss them at Florence, Italy!