Compare to non-neural baseline

[jbingel]

Good news: with our simple features (word length, frequency, character perplexity, target/sentence similarity), random forest gets F1=0.69 on the German data, compared to roughly 0.75 with the best neural MTL model

[jbingel]

For English and Spanish, however, random forest seems to strongly outperform neural models:

	neural	RF
EN	0.70	0.78
ES	0.64	0.69
DE	0.75	0.69

[jbingel]

new features improved neural models:

	neural	RF
EN	0.74	0.79
ES	0.69	0.69
DE	0.76	0.69

[bjerva]

Nice!!!

[jbingel]

Yeah. For the final submission (especially for EN and ES) we could try an ensemble between neural and RF (and others...). I'll look to get this implemented.