Digital Marketing Dhaka — DMD agency logo
Call
Bilingual NLU · Dhaka

Bangla + English NLU for Chatbots in Dhaka

ঢাকায় বাংলা + ইংরেজি NLU চ্যাটবট

Bangla + English NLU for Chatbots in Dhaka — Digital Marketing Dhaka caricature illustration
Banglish ready
মিশ্র ভাষা

One bot that actually understands 'দাম কত?', 'daam koto?', 'what's the price?' and everything in between — without misfiring or defaulting to English.

একটাই bot যা 'দাম কত?', 'daam koto?', 'price koto?' সবকিছু বোঝে — misfire না করে, ইংরেজিতে default না গিয়ে।

Call to discussGet a proposal
Quick answer

Bangla + English NLU — in one paragraph

Bangla + English NLU (Natural Language Understanding) for Dhaka chatbots means the bot correctly detects intent whether the customer types in Bangla script (দাম কত?), Roman-Bangla transliteration (daam koto?) or English — and routes to the right answer or human agent. Digital Marketing Dhaka ships bilingual NLU using Gemini embeddings + hybrid retrieval so a single conversation can flow between all three writing styles without breaking context.

ঢাকার চ্যাটবটের জন্য বাংলা + ইংরেজি NLU মানে — কাস্টমার বাংলা স্ক্রিপ্ট, Roman-Bangla (daam koto?) বা ইংরেজি — যেভাবেই লিখুক, বট সঠিক intent detect করে সঠিক উত্তর বা human agent-এ route করে।

Why Dhaka

Why this matters for Bangladeshi businesses

Dhaka customers switch languages mid-sentence. A typical WhatsApp chat: 'Bhai eta ki available? size L in blue? দাম কত?'. Most off-the-shelf chatbots misfire — either the Bangla script breaks their model, or 'daam' isn't in their intent list. Ours handles all three writing systems as one intent because we use multilingual embeddings + a hybrid keyword layer that maps 'daam' → 'দাম' → 'price'. This is the difference between a bot that converts and a bot that gets screenshotted for a meme.

ঢাকার কাস্টমার এক sentence-এ language switch করে। 'Bhai eta ki available? size L in blue? দাম কত?' — বেশিরভাগ চ্যাটবট এখানে misfire করে। আমাদের NLU multilingual embedding + hybrid keyword layer ব্যবহার করে যেটা 'daam' → 'দাম' → 'price' map করে।

What's included

Every piece we ship in this engagement

01

Multilingual embeddings

Gemini embedding-001 or OpenAI text-embedding-3-large — trained on Bangla, English and enough Banglish that transliteration doesn't need a separate pass.

02

Intent taxonomy in Bangla + English

Every intent (price, stock, EMI, delivery, refund, complaint) documented in both languages with 15–25 example utterances covering script, transliteration and mixed.

03

Hybrid retrieval (BM25 + vector)

Keyword layer catches brand names, product codes and numbers that embeddings miss; vector layer catches paraphrasing. Combined = 15–25% better recall on Banglish.

04

Transliteration handling

'daam koto', 'daam koto', 'দাম কত' — all mapped to the same intent. We use Avro/Google Bangla transliteration tables + real chat logs from your team.

05

Confidence thresholds + fallback

Every reply has a confidence score. Below threshold → 'let me connect you with our team' + WhatsApp handoff. Threshold tuned to your risk tolerance.

06

Language-preserving replies

Bot replies in the language the user wrote in. Bangla in → Bangla out. Banglish in → Banglish out (not corrected). This is what makes it feel human, not a machine translation.

07

Slang and dialect coverage

We seed the model with local Dhaka slang — 'ostad', 'boss', 'bhai', 'apu', 'nishchoy', 'obosshoi' — so casual customers aren't treated like broken input.

08

Monthly intent gap review

Every month we surface the top 20 misfired intents and add them to the taxonomy. Bot accuracy goes from 78% at launch → 92%+ by month 3.

Our process

How we deliver it in Dhaka

01

Intent workshop (Week 1)

We audit 200+ real chat logs from your team and design a bilingual intent taxonomy signed off with sales + support.

02

Embedding + retrieval (Week 2)

Vector store built with Bangla + English data, hybrid retrieval tuned against 100 test queries in all three writing styles.

03

Supervised launch (Weeks 3–4)

Live with shadow mode — your team reviews every bot reply for language accuracy before send. Threshold auto-tightens as accuracy climbs.

04

Monthly retraining

Misfired conversations added to the corpus, intent gaps closed, transliteration edge-cases patched. Accuracy compounds every month.

Deep dive

The details that decide the outcome

Why English-only models break in Dhaka

OpenAI's text-embedding-3-small was trained on ~93% English data. Bangla script queries get lower-quality embeddings, which means the retrieval layer misses relevant chunks. Gemini's multilingual model is meaningfully better on Bangla (indic languages were a training priority) — in our benchmarks on 500 Dhaka chat logs, Gemini beats OpenAI on Bangla recall by 22%. This is why we default to Gemini for Bangla-heavy deployments and only switch to GPT for English-heavy B2B SaaS.

Banglish is not a translation problem

Naive systems try to translate 'daam koto?' → 'what is the price?' → run through English model. This loses tone, context and speed. Our approach: treat Banglish as a first-class language. Multilingual embeddings + a small transliteration alias table means 'daam' and 'দাম' resolve to the same intent vector in one hop. No round-trip translation, no lost context, sub-second response.

The hybrid retrieval trick

Pure embeddings miss exact-match queries like 'stock ache SKU-4471?'. Pure BM25 keyword miss paraphrases like 'this available?'. Hybrid retrieval runs both, weights the results (typically 0.7 vector + 0.3 keyword), and returns the union. On our Dhaka benchmark set (500 real WhatsApp queries), hybrid beat pure-vector by 18% top-1 accuracy — the single biggest quality lever in the whole pipeline.

How we prove the NLU works before launch

Every bot ships with a public accuracy dashboard: intent taxonomy on one axis, 100 test queries on the other, colour-coded pass/fail. You see, before launch, that 'ei product ki EMI te ache?' correctly resolves to intent = EMI_options with 94% confidence. No black-box magic — just numbers your ops team can audit.

Dhaka proof

What it looks like on the ground

A private hospital in Dhanmondi launched a bilingual appointment bot on WhatsApp. Initial vendor (English-only) had 61% intent accuracy and patients frequently rage-quit to the operator. We rebuilt on Gemini with hybrid retrieval and a 32-intent bilingual taxonomy. Post-launch: 89% intent accuracy on Bangla script, 91% on Banglish, 94% on English. Human hand-off rate dropped from 47% → 18%, and the front-desk team reallocated one FTE to outbound follow-up. 4-month recorded ROI: BDT 6.4 lakh in saved operator time.

ধানমন্ডির একটা private hospital WhatsApp-এ bilingual appointment bot launch করল। আগের vendor (English-only) 61% intent accuracy — patients operator-এ চলে যেত। আমরা Gemini + hybrid retrieval + 32-intent bilingual taxonomy দিয়ে rebuild করলাম। Post-launch: বাংলা script-এ 89%, Banglish-এ 91%, English-এ 94% accuracy। Human handoff rate 47% → 18%। ৪ মাসে saved operator time-এ BDT ৬.৪ লাখ ROI।

FAQ

Frequently asked questions

Do you need Bangla training data from us?

Ideally yes — 200+ real chat logs give the best accuracy. If you don't have them, we bootstrap with synthetic + public Bangla datasets and retrain after month 1 with real conversations.

Can the bot switch languages mid-reply?

Yes — if the user writes 'price koto? and delivery time?', the bot answers in matching Banglish. Language mirroring is on by default.

What accuracy should we expect?

78–85% at launch, 90%+ by month 3 with monthly retraining. Anything claiming '99% accuracy' out of the box for Bangla is marketing, not measurement.

Does it handle voice notes?

Yes — WhatsApp voice notes transcribed with Gemini's Bangla ASR, then passed through the same NLU. Accuracy is slightly lower than text (~80%) so we default to sending the transcript back for user confirmation.

আমাদের কাছ থেকে বাংলা training data লাগবে?

আদর্শভাবে হ্যাঁ — 200+ real chat log সবচেয়ে ভালো accuracy দেয়। না থাকলে synthetic + public dataset দিয়ে শুরু, ১ মাস পর real conversation দিয়ে retrain।

বট কি মাঝপথে ভাষা বদলাতে পারে?

হ্যাঁ — user 'price koto? and delivery time?' লিখলে বট Banglish-এ উত্তর দেয়। Language mirroring default on।

কেমন accuracy expect করব?

Launch-এ 78-85%, monthly retraining সহ মাসে 3-এ 90%+। বাংলার জন্য '99% accuracy out of the box' claim marketing, measurement নয়।

Voice note handle করে?

হ্যাঁ — WhatsApp voice note Gemini Bangla ASR দিয়ে transcribe, তারপর same NLU-তে pass। Accuracy ~80%।

Related AI deliverables

Keep exploring the AI engagement

01

Custom LLM chatbot

ঢাকায় কাস্টম LLM চ্যাটবট ডেভেলপমেন্ট

A GPT-style assistant trained on your own catalogue, FAQs and policies — deployed on your site and WhatsApp in 3–4 weeks with proper RAG grounding, not a generic ChatGPT wrapper.

Read more →
05

WhatsApp AI concierge

ঢাকার ব্র্যান্ডদের জন্য WhatsApp AI কনসিয়ার্জ

A 24/7 AI concierge on your WhatsApp Business number — answers in Bangla + English, shares your catalogue, sends payment links, handles COD, and hands off to sales when it matters.

Read more →
04

Voice search SEO

ঢাকার ব্যবসার জন্য Voice Search SEO

Own the 'near me' voice query. FAQ schema, SpeakableSpecification and conversational Bangla + English content so Google Assistant reads your answer aloud — first.

Read more →
Free 30-min consult

Talk to a Dhaka specialist about bangla + english nlu

Share your number — a senior strategist will call you back within one business day with a short, honest audit and next steps.

No spamNo obligationReply in 24h
Service: Bangla + English NLU

No spam. Reply within 1 business day.

Ready to ship this in your next 30 days?

Book a free 30-min AI strategy call. We'll audit your current chatbot / GEO / voice setup live and show you the biggest wins — no obligation.

Call +880 1876-372190Request a proposal