Task-oriented dialogue understanding remains underexplored for Bangla, particularly in multilingual settings involving English and romanized Bangla (Banglish). We introduce BanglaTOD, a parallel tri-lingual benchmark comprising 209 multi-turn dialogues across nine Bangladesh-specific service domains and 64 intents. Each conversational turn is aligned across Bangla, English, and Banglish, yielding 4,689 language-specific turn records annotated with intents, token-level slots, state updates, and accumulated dialogue states. We establish context-aware baselines for joint intent classification and slot filling using mBERT, BERT-base-uncased, and BanglaBERT. Cross-language analysis reveals substantial performance differences across models and linguistic representations, while removing dialogue history reduces intent accuracy by 19.0--25.7 percentage points. BanglaTOD provides a foundation for multilingual dialogue understanding and future dialogue state tracking research in Bangladesh-specific conversational applications. The dataset is publicly available at Hugging Face.
Figures & tables
Domain
Dialogues
Intents
Education
50
9
Government Service
32
19
Healthcare
25
5
Mobile Financial Serv.
25
5
E-commerce
25
10
Restaurant
16
5
Table 1: Domain and intent coverage of the dataset.
Characteristic
Value
Total dialogues
209
Languages
3
Domains
9
Intents
64
Turns per language
1,563
Language-specific turn records
4,689
Table 2: Overview of BanglaTOD.
Language
Tokens
Vocabulary
Avg. Tokens/Turn
Bangla
15,726
2,448
10.06
English
17,858
2,192
11.43
Banglish
15,412
2,661
9.86
Table 3: Lexical statistics for the language variants.
Intent / Domain
Slot (entity-level)
Joint
Model / Language
Dom. Acc
Micro-F1
Macro-F1
Wtd-F1
Micro-F1
Macro-F1
Wtd-F1
EM
mBERT
Overall
0.859 ±.007
0.601 ±.012
0.408 ±.015
0.623 ±.012
0.330 ±.003
0.194 ±.003
0.302 ±.006
0.219 ±.010
Bangla
0.869 ±.018
0.605 ±.019
0.453 ±.007
0.613 ±.019
0.323 ±.011
0.203 ±.003
0.285 ±.015
0.225 ±.009
English
0.859 ±.006
0.625 ±.007
0.482 ±.009
0.636 ±.010
0.345 ±.006
0.231 ±.009
0.318 ±.004
0.237 ±.006
Banglish
0.849 ±.007
0.573 ±.013
0.426 ±.014
0.592 ±.008
0.321 ±.005
0.212 ±.004
0.301 ±.004
0.195 ±.018
Table 4: Test results as mean ±std over seeds 42, 43, 44 (context-aware input). Domain accuracy is derived from the predicted intent. Dom. = domain, Wtd = weighted, EM = exact match, Std = standard deviation.
Model
Metric
Ctx
No Ctx
Δ
mBERT
Intent acc.
0.605
0.356
+0.249
Token macro-F1
0.250
0.215
+0.035
BERT-base.
Intent acc.
0.578
0.321
+0.257
Token macro-F1
0.216
0.179
+0.037
BanglaBERT
Intent acc.
0.385
0.195
+0.190
Token macro-F1
0.128
0.106
+0.021
Table 5: Effect of dialogue context. Ctx = up to three previous turns; Δ = Ctx − No ctx.
Intent
mBERT
BERT
BanglaBERT
hotel_book_cancel
1.000
1.000
0.882
nid_card_service
0.250
0.154
0.000
police_services
0.000
0.154
0.267
govt_fin_assist
0.333
0.231
0.100
land_records
0.588
0.286
0.190
tax_services
0.000
0.476
0.286
Table 6: Selected per-intent F1 scores on the test set at the seed-42 checkpoint.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Value
Optimizer
AdamW ( Loshchilov and Hutter, 2019 )
Learning rate
1.5×10−5
LLRD
0.95
Weight decay
0.01
Scheduler
Cosine, 10% warm-up
Batch size
16
Appendix
Table 7: Shared training configuration.
Error type
Slot label
P
R
Support
Reliable
B-booking_id
1.000
1.000
9
B-date_of_birth
1.000
1.000
3
Omitted
B-action
1.000
0.500
6
B-hotel_name
1.000
0.333
3
B-appl_status
0.000
0.000
3
Low precision
B-query
0.208
0.333
15
Appendix
Table 8: Representative token-level slot predictions for mBERT at the seed-42 checkpoint. P = precision; R = recall; support = number of tokens.
State tracking studies how AI systems maintain and update information about entities, tasks, or environments across sequential observations and actions.