Table 2.

Results on recreating known lexicons.

(a) Corpus methods outperform WordNet on standard English. Using word-vector embeddings learned on a massive corpus (1011 tokens), we see that both corpus-based methods outperform the WordNet-based approach overall.
Method	AUC	Ternary F1	τ
SentProp	90.6	58.6	0.44
Densifier	93.3	62.1	0.50
WordNet	89.5	58.7	0.34
Majority	–	24.8	–

(b) Corpus approaches are competitive with a distantly supervised method on Twitter. Using Twitter embeddings learned from ~109 tokens, we see that the semi-supervised corpus approaches using small seed sets perform very well.
Method	AUC	Ternary F1	τ
SentProp	86.0	60.1	0.50
Densifier	90.1	59.4	0.57
Sentiment140	86.2	57.7	0.51
Majority	–	24.9	–

(c) SentProp performs best with domain-specific finance embeddings. Using embeddings learned from financial corpus (~2× 107 tokens), SentProp significantly outperforms the other methods.
Method	AUC	Ternary F1
SentProp	91.6	63.1
Densifier	80.2	50.3
PMI	86.1	49.8
CountVecs	81.6	51.1
Majority	–	23.6

(d) SentProp performs well on standard English even with 1000x reduction in corpus size. SentProp maintains strong performance even when using embeddings learned from the 2000s decade of COHA (only 2 × ~107 tokens).
Method	AUC	Ternary F1	τ
SentProp	83.8	53.0	0.28
Densifier	77.4	46.6	0.19
PMI	70.6	41.9	0.16
CountVecs	52.7	32.9	0.01
Majority	–	24.3	–