Table 6.
Test set and model | About suicide (n=478) | Off-topic (n=163) | |||||
|
Precision (95% CI) | Recall (95% CI) | F 1 | Precision (95% CI) | Recall (95% CI) | F 1 | |
TF-IDFa and SVMb | 0.89 (85.74-91.71) | 0.85 (80.96-87.64) | 0.87 | 0.60 (53.03-67.49) | 0.69 (61.63-76.30) | 0.65 | |
BERTc,d | 0.90 (87.42-92.81) | 0.94 (91.64-96.07) | 0.92 | 0.80 (71.62-85.67) | 0.68 (60.35-75.17) | 0.73 | |
XLNetd | 0.90 (87.12-92.59) | 0.93 (90.68-95.38) | 0.92 | 0.76 (68.60-83.06) | 0.67 (59.72-74.60) | 0.71 |
aTF-IDF: term frequency-inverse document frequency.
bSVM: support vector machine.
cBERT: Bidirectional Encoder Representations from Transformers.
dScores are averages across 5 model runs for BERT and XLNet. Table S5 in Multimedia Appendix 1 shows separate runs.