Fully labeled parse trees (this is based upon Pang and Lee's 2005 "sentence-polarity dataset")
Every node in the trees are labelled according to the sentiment of the phrase it represents:
Adapted from Socher et al 2013b
| Error rate | Strategy | Reported by | Notes |
|---|---|---|---|
| 20.6% | SVMs | Socher et al., 2013b | |
| 19.9% | Word Vector averaging | Socher et al., 2013b | |
| 18.2% | Naïve Bayes | Socher et al., 2013b | |
| 17.6% | Recursive Neural Nets | Socher et al., 2013b | |
| 17.1% | Matrix Vector-RNN | Socher et al., 2013b | |
| 16.9% | Bi-gram Naïve Bayes | Socher et al., 2013b | |
| 14.6% | Recursive Neural Tensor Network | Socher et al., 2013b | |
| 12.2% | Paragraph Vector | Le and Mikolov 2014 | Using a mixture of PV-DM and PV-DBOW |
| Error rate | Strategy | Reported by | Notes |
|---|---|---|---|
| 67.3% | Word Vector averaging | Socher et al., 2013b | |
| 59.3% | SVMs | Socher et al., 2013b | |
| 59.0% | Naïve Bayes | Socher et al., 2013b | |
| 58.1% | Bi-gram Naïve Bayes | Socher et al., 2013b | |
| 56.8% | Recursive Neural Net | Socher et al., 2013b | |
| 55.6% | Matrix Vector-RNN | Socher et al., 2013b | |
| 54.3% | Recursive Neural Tensor Network | Socher et al., 2013b | |
| 51.3% | Paragraph Vector | Le and Mikolov 2014 | Using a mixture of PV-DM and PV-DBOW |
-
Le and Mikolov 2014: Distributed Representations of Sentences and Documents
-
Socher et al., 2013a: Reasoning with neural tensor networks for knowledge base completion.
-
Socher et al., 2013b : Recursive deep models for semantic compositionality over a sentiment treebank.