group-data-ai/transaction-narrative-ner
Transaction Narrative Entity Extractor
Entity F1 (micro)
0.912
Counterparty name F1
0.938
Purpose classification accuracy
0.864
Batch throughput
9,000 rows/s
About this model
Pulls structured entities out of free-text payment narratives on DuitNow, IBG and FPX transfers: merchant or counterparty name, payment purpose (rent, salary, loan repayment, gift, investment), reference numbers and platform names. Handles English, Bahasa Malaysia and typical abbreviations ("sewa", "gaji", "byr").
Intended use
Feature enrichment for fraud, mule and AML models, spend categorisation, and investigator search. Output enriches transactions; it does not decide anything on its own.
Training data lineage
Token-classification model fine-tuned from a multilingual encoder on 120,000 narratives sampled from casa-transactions in BigQuery and labelled in-house by Financial Crime analysts. Trained, evaluated and registered in Vertex AI; batch inference runs nightly in BigQuery through a remote model.
Limitations & bias notes
Narratives under three characters or made only of emojis carry no signal. Purpose labels are indicative; "gift" and "loan" are often confused in informal transfers between family members.
Evaluation metrics
| Entity F1 (micro) | 0.912 |
| Counterparty name F1 | 0.938 |
| Purpose classification accuracy | 0.864 |
| Batch throughput | 9,000 rows/s |
Try it
Live sandboxFeedback
Owning business unit
- Updated
- 2026-07-06
- Latest version
- 1.3.0
- Licence
- Group Reuse
- Access
- Open to all BUs
Trained on
casa-transactions →