Table 4.

The effectiveness of different CRF feature groups; all feature set (All) includes: context, lexicon, POS, negation, and embedding clusters (cluster)

CRF Features	DS			Twitter
	P	R	F	P	R	F
All	0.856	0.776	0.814	0.765	0.682	0.721
All – lexicon	0.852	0.781	0.815	0.765	0.646	0.701
All – POS	0.853	0.776	0.812	0.754	0.653	0.700
All – negation	0.854	0.769	0.810	0.752	0.646	0.695*
All – context	0.811	0.665	0.731*	0.624	0.498	0.554*
All – cluster	0.851	0.745	0.794*	0.788	0.549	0.647*
Context + cluster	0.860	0.784	0.821*	0.746	0.628	0.682*

Statistically significant changes (p < 0.05), when compared with All feature set, are marked with asterisks.

The highest values in each column are highlighted in bold.