Skip to main content
. Author manuscript; available in PMC: 2025 Mar 5.
Published in final edited form as: Proc Int AAAI Conf Weblogs Soc Media. 2024 May 28;18:422–434. doi: 10.1609/icwsm.v18i1.31324

Table 5:

Observations with the Out-of-distribution (OOD) dataset. Comparison of the explanations obtained by LIME with three different types of textual cues annotated as Trigger, LoST indicators, and consequences using ROUGE scores and BLEU scores. (italicize and underline) represents the highest values of LoST in True Positives and comparable values of LoST and Trigger, respectively.

Model Eval. T L C

BERT ROUGE 0.0449 0.1764 0.0533
BLEU 0.0225 0.1328 0.0432
ALBERT ROUGE 0.1525 0.1857 0.0259
BLEU 0.1142 0.1468 0.0166
DistilBERT ROUGE 0.0569 0.1291 0.0266
BLEU 0.0138 0.0993 0.0200
DeBERTa ROUGE 0.0979 0.1094 0.0159
BLEU 0.0638 0.0739 0.0100

ClinicalBERT ROUGE 0.0607 0.0968 0.0533
BLEU 0.0279 0.0536 0.0500
PsychBERT ROUGE 0.1444 0.2310 0.0352
BLEU 0.1018 0.1843 0.0260
MentalBERT ROUGE 0.1054 0.2228 0.0214
BLEU 0.0633 0.1897 0.0133