This is from the "Conclusions" slide:
• Deep recurrent nets perform better than their shallow
counterparts of the same size on both DSE and ESE
extraction.
• Both shallow and deep RNNs capture aspects of
subjectivity, but deep RNNs seem to better handle the
expression boundaries.
• Deep RNNs outperforms previous baselines CRF and
semi-CRF without having access to the dependency or
constituency trees, opinion lexicons or POS tags, even
when (semi)CRF has access to word vectors.
tldr: Deep RNN's are better than shallow RNN's and CRF's for several NLP tasks.