ACSA-Clothing v4 Demo

Overall sentiment baselines and per-aspect Proposed vs no-metadata comparison.

Baseline 1 TFIDF LogReg

F1 0.6364

Acc 0.8228

Baseline 2 BERT overall 3class

F1 0.7465

Acc 0.9072

Baseline 3 BERT ACSA no meta / aggregated to overall

F1 0.4803

Acc 0.7402

Proposed BERT Meta Fusion / overall head

F1 0.7308

Acc 0.9033

Proposed BERT Meta Fusion / aggregated to overall (reference)

F1 0.4883

Acc 0.7487

Overall 3-class comparison

Overall 3-class comparison
Model	Macro-F1	Accuracy	Macro-F1 (%)	Accuracy (%)
Proposed BERT Meta Fusion / aggregated to overall (reference)	0.6364	0.8228	63.64%	82.28%

Model	Macro-F1	Accuracy	Macro-F1 (%)	Accuracy (%)
Baseline 1 TFIDF LogReg	0.6364	0.8228	63.64%	82.28%
Baseline 2 BERT overall 3class	0.7465	0.9072	74.65%	90.72%
Baseline 3 BERT ACSA no meta / aggregated to overall	0.4803	0.7402	48.03%	74.02%
Proposed BERT Meta Fusion / overall head	0.7308	0.9033	73.08%	90.33%
Proposed BERT Meta Fusion / aggregated to overall (reference)	0.4883	0.7487	48.83%	74.87%

Per-aspect comparison

Per-aspect comparison
Aspect	Proposed F1	No-Meta F1	F1 Gain	Proposed Acc	No-Meta Acc	Acc Gain
APPEARANCE	0.8077	0.7458	0.0619	0.8831	0.8577	0.0255

Aspect	Proposed F1	No-Meta F1	F1 Gain	Proposed Acc	No-Meta Acc	Acc Gain
SIZE	0.8077	0.7458	0.0619	0.8831	0.8577	0.0255
MATERIAL	0.7759	0.7433	0.0326	0.8586	0.8457	0.0129
QUALITY	0.7832	0.757	0.0262	0.864	0.853	0.011
APPEARANCE	0.8071	0.7627	0.0444	0.8871	0.8676	0.0195
STYLE	0.7983	0.7551	0.0432	0.8767	0.8541	0.0226
VALUE	0.7975	0.7572	0.0403	0.8749	0.854	0.0209

A1 removes text metadata, A2 removes numeric metadata, A3 replaces Cross-Attention with concat fusion.

Proposed

F1 0.7949

Acc 0.8741

A1_no_text_meta

F1 0.7574

Acc 0.8524

A2_no_numeric_meta

F1 0.7956

Acc 0.8735

A3_concat_fusion

F1 0.7961

Acc 0.8764

Ablation summary

Ablation summary
Variant	Mean Macro-F1	Mean Accuracy	Macro-F1 (%)	Accuracy (%)
A2_no_numeric_meta	0.7949	0.8741	79.49%	87.41%

Variant	Mean Macro-F1	Mean Accuracy	Macro-F1 (%)	Accuracy (%)
Proposed	0.7949	0.8741	79.49%	87.41%
A1_no_text_meta	0.7574	0.8524	75.74%	85.24%
A2_no_numeric_meta	0.7956	0.8735	79.56%	87.35%
A3_concat_fusion	0.7961	0.8764	79.61%	87.64%

Per-aspect ablation details

Per-aspect ablation details
Aspect	Proposed F1	Proposed Acc	A1_no_text_meta F1	A1_no_text_meta Acc	A1_no_text_meta F1 Delta	A2_no_numeric_meta F1	A2_no_numeric_meta Acc	A2_no_numeric_meta F1 Delta	A3_concat_fusion F1	A3_concat_fusion Acc	A3_concat_fusion F1 Delta
APPEARANCE	0.8077	0.8831	0.7545	0.8554	-0.0531	0.8051	0.8798	-0.0026	0.8053	0.8836	-0.0023

Aspect	Proposed F1	Proposed Acc	A1_no_text_meta F1	A1_no_text_meta Acc	A1_no_text_meta F1 Delta	A2_no_numeric_meta F1	A2_no_numeric_meta Acc	A2_no_numeric_meta F1 Delta	A3_concat_fusion F1	A3_concat_fusion Acc	A3_concat_fusion F1 Delta
SIZE	0.8077	0.8831	0.7545	0.8554	-0.0531	0.8051	0.8798	-0.0026	0.8053	0.8836	-0.0023
MATERIAL	0.7759	0.8586	0.7461	0.84	-0.0298	0.7766	0.8591	0.0008	0.785	0.8665	0.0091
QUALITY	0.7832	0.864	0.7587	0.8481	-0.0245	0.7887	0.8651	0.0055	0.7934	0.8719	0.0102
APPEARANCE	0.8071	0.8871	0.7645	0.864	-0.0426	0.8025	0.8842	-0.0046	0.8024	0.8851	-0.0047
STYLE	0.7983	0.8767	0.7617	0.8556	-0.0365	0.8047	0.8793	0.0065	0.7935	0.8741	-0.0047
VALUE	0.7975	0.8749	0.7592	0.8513	-0.0383	0.7962	0.8733	-0.0013	0.7972	0.8773	-0.0002

Fine-grained explanation for enterprise diagnostics: aspect predictions, metadata attention, and operational recommendations.

Review text

features_text

categories_text

price

average_rating

rating_number

Aspect-level predictions

Aspect-level predictions
1	2	3

Identify the good/bad-review keywords behind the model result for consumer shopping reference.

Review text

features_text

categories_text

price

average_rating

rating_number

Keyword evidence

Keyword evidence
1	2	3

Consumer shopping reference

ACSA-Clothing v4: BERT + Metadata Cross-Attention