The performance of BERT as data representation of text clustering

Subakti, Alvin; Murfi, Hendri; Hariadi, Nora

doi:10.1186/s40537-022-00564-9

Journal of Big Data

Table 4 Cluster evaluation on R2 dataset

From: The performance of BERT as data representation of text clustering

Method	AG news
Method	ACC	NMI	ARI
TFIDF + KM	0.8471 ± 0	0.5034 ± 0	0.5033 ± 0
BERT + Max + I + KM	0.8457 ± 0	0.5025 ± 0	0.5024 ± 0
BERT + Max + LN + KM	0.8472 ± 0	0.5052 ± 0	0.5052 ± 0
BERT + Max + N + KM	0.8469 ± 0	0.4985 ± 0.0015	0.4984 ± 0.0015
BERT + Max + MM + KM	0.8495 ± 0	0.4942 ± 0	0.4941 ± 0
BERT + Mean + I + KM	0.8471 ± 0	0.5034 ± 0	0.5033 ± 0
BERT + Mean + LN + KM	0.8507 ± 0	0.5036 ± 0	0.5035 ± 0
BERT + Mean + N + KM	0.8507 ± 0	0.5036 ± 0	0.5035 ± 0
BERT + Mean + MM + KM	0.6624 ± 0.0002	0.0822 ± 0.0003	0.0821 ± 0.0003
TFIDF + EFCM	0.8476 ± 0	0.5043 ± 0	0.5042 ± 0
BERT + Max + I + EFCM	0.8462 ± 0	0.5034 ± 0	0.5033 ± 0
BERT + Max + LN + EFCM	0.8474 ± 0	0.504 ± 0	0.5039 ± 0
BERT + Max + N + EFCM	0.8479 ± 0	0.4964 ± 0	0.4964 ± 0
BERT + Max + MM + EFCM	0.8498 ± 0	0.4957 ± 0	0.4957 ± 0
BERT + Mean + I + EFCM	0.8476 ± 0	0.5043 ± 0	0.5042 ± 0
BERT + Mean + LN + EFCM	0.8505 ± 0	0.5 ± 0	0.4999 ± 0
BERT + Mean + N + EFCM	0.8505 ± 0	0.5 ± 0	0.4999 ± 0
BERT + Mean + MM + EFCM	0.6636 ± 0	0.0827 ± 0	0.0826 ± 0
TFIDF + DEC	0.859 ± 0.0100	0.5064 ± 0.0205	0.5158 ± 0.0288
BERT + Max + I + DEC	0.793 ± 0.0794	0.386 ± 0.1525	0.3545 ± 0.1835
BERT + Max + LN + DEC	0.8409 ± 0.0188	0.4827 ± 0.0308	0.466 ± 0.0480
BERT + Max + N + DEC	0.8474 ± 0.0033	0.4996 ± 0.0078	0.4825 ± 0.0092
BERT + Max + MM + DEC	0.7816 ± 0.0590	0.3727 ± 0.1332	0.3269 ± 0.1348
BERT + Mean + I + DEC	0.8497 ± 0.0025	0.504 ± 0.0068	0.4891 ± 0.0070
BERT + Mean + LN + DEC	0.8494 ± 0.0017	0.5035 ± 0.0059	0.4882 ± 0.0047
BERT + Mean + N + DEC	0.8533 ± 0.0045	0.4996 ± 0.0059	0.4993 ± 0.0128
BERT + Mean + MM + DEC	0.6373 ± 0	0.00002 ± 0.0001	0.00001 ± 0.00009
TFIDF + IDEC	0.8654 ± 0.0116	0.5213 ± 0.0252	0.5345 ± 0.0342
BERT + Max + I + IDEC	0.8095 ± 0.0616	0.4303 ± 0.1373	0.3917 ± 0.1442
BERT + Max + LN + IDEC	0.8401 ± 0.0228	0.485 ± 0.0349	0.4643 ± 0.0572
BERT + Max + N + IDEC	0.8428 ± 0.0297	0.4889 ± 0.0704	0.4718 ± 0.0686
BERT + Max + MM + IDEC	0.7815 ± 0.0588	0.3623 ± 0.1399	0.3255 ± 0.1366
BERT + Mean + I + IDEC	0.8494 ± 0.0011	0.507 ± 0.0049	0.4881 ± 0.0032
BERT + Mean + LN + IDEC	0.8494 ± 0.0007	0.5045 ± 0.0048	0.4884 ± 0.0021
BERT + Mean + N + IDEC	0.8518 ± 0.0038	0.4952 ± 0.0068	0.4951 ± 0.0108
BERT + Mean + MM + IDEC	0.6374 ± 0.0002	0.0007 ± 0.0014	0.0004 ± 0.0007

The feature extraction and normalization strategies are abbreviated into Max for max pooling, Mean for mean pooling, I for identity normalization, LN for layer normalization, N for standard normalization, and MM for min–max normalization.The deviations denote the standard deviation of the metric from 50 repetitions. The values in bold denote the highest value in every metric in each text clustering algorithm. While the method in bold, if there are any, is the best performing method in each text clustering algorithm

Back to article page