Intent Detection for Chatbots: Taxonomy, Model, Monitoring

Intent Detection for Chatbots: Taxonomy, Model, Monitoring

"I want to order pizza", "status of my order", "how to get a refund" — three different intents. If the chatbot cannot distinguish them, the user leaves. Without quality intent detection, a bot cannot correctly route requests, handle errors, or sustain a dialogue. Misclassification leads to customer loss and increased operator load. We solve this by designing the taxonomy, selecting the model (BERT, LLM, or hybrid approach), and implementing monitoring.

The result: >90% accuracy on target intents, p99 latency <200 ms. Our intent detection solutions have been tested under real loads: up to 10,000 requests per day, with quality degradation of no more than 3% after six months. Experience: 7+ years in NLP, 20+ projects, including bots for e-commerce, fintech, and technical support. Savings on request handling for clients: up to 60% reduction in manual support costs. Fast payback — within 2–3 months. Typical project cost ranges from $5,000 to $20,000 depending on complexity.

Designing Intent Taxonomy

Rule: each intent is a single concrete intention with a clear handler. Common mistakes:

• Too broad intent: help — unclear what to do.
• Duplication: order_status and check_my_order — same intent.
• Fuzzy boundaries: complaint vs negative_feedback — often ambiguous.

For a moderately complex chatbot: 20–50 intents. More than 100 is a sign of poor architecture; revise the hierarchy.

Why Hybrid Approach Is Better Than Classic

The classic approach (Rasa NLU, Dialogflow) trains on labeled examples: deterministic, fast (<10 ms), but requires 50–200 examples per intent and does not adapt without retraining. LLM-based (GPT-4o-mini with few-shot) requires no training, is flexible, but slower (200–500 ms) and more expensive in cost per token. Hybrid — BERT as the first pass (fast, cheap), LLM as fallback for low-confidence and edge cases. In our projects, we use exactly this scheme: it provides a balance of speed and coverage.

Comparison of key metrics: hybrid is 10x faster than pure LLM with comparable coverage. More about the BERT model.

How We Implement the Hybrid Classifier

1. Audit and taxonomy design. Analyze dialogue history, identify frequent topics, build an intent map. Conduct a workshop with your team to define intent boundaries. Result: a document with hierarchy, slots, and fallback rules.

2. Dataset labeling. Collect 100–300 examples per intent (real user phrases). Add confusion matrix negative examples — phrases from similar intents to "stretch" class boundaries. Example: "I want to change my order" vs "I want to cancel my order" — different intents, but both must be in the dataset.

3. Training and calibration. Fine-tune BERT (e.g., DeepPavlov/rubert-base-cased) with cross-entropy. Calibrate confidence threshold (usually 0.6): when confidence is lower, pass the request to LLM.

from transformers import pipeline

intent_classifier = pipeline(
    "text-classification",
    model="./intent_classifier",
    tokenizer="DeepPavlov/rubert-base-cased",
    top_k=3
)

def detect_intent(text: str) -> IntentResult:
    results = intent_classifier(text)
    top = results[0]

    if top["score"] < 0.6:
        # Fallback to LLM
        return llm_classify_intent(text)

    return IntentResult(intent=top["label"], confidence=top["score"])

4. Integration and deployment. Deploy the model via Triton Inference Server or ONNX Runtime — latency p99 <50 ms on GPU. Set up monitoring: confusion matrix weekly, OOS logs, feedback pipeline. We ensure the model does not degrade on new data: if accuracy drops below 85%, an automatic trigger initiates retraining.

Monitoring Confusion Matrix and Preventing Drift

The confusion matrix is the main tool for identifying problematic intent pairs. If, for example, order_status and change_order are often confused, add more confusing negatives to the dataset. Regular analysis of OOS logs helps detect new user intentions that should be extracted as separate intents. As a result, classification accuracy does not decline over time.

Project Deliverables

• Documentation of intent taxonomy (Google Docs / Confluence)
• Labeled dataset (JSONL, CSV format)
• Trained model + Docker image for deployment
• Integration with your bot (REST API, gRPC)
• Load testing (result: p99 latency <200 ms)
• Monitoring and alerting (Grafana dashboards)
• Training of two of your engineers on the pipeline

Timeline and How to Order

Development timeline: from 2 to 6 weeks depending on taxonomy complexity and dataset size. Cost is calculated individually after an audit. We will evaluate your project for free — contact us for a consultation. Order intent detection development and reduce interpretation errors by 30–70%. Get a consultation — we will select the optimal stack for your budget. Our team is certified in MLOps and guarantees performance improvements.

NLP Development: Text Classification, NER, Embeddings, and Information Extraction

We often receive a task: process 50,000 support tickets — currently all manual. Dataset — 3,000 labeled examples, 12 categories, imbalance: one category occupies 40% of the sample, three at 1-2% each. Baseline accuracy — 78%. Sounds decent until you look at recall for rare classes: 0.31, 0.44, 0.28. These classes — complaints and churn threats — are most important to the business.

This is a typical NLP development project. The problem is not the algorithm but that accuracy is the wrong metric. Our experience across 30+ projects shows: we start by analyzing business metrics and only then choose the model.

Why accuracy is not the right metric for rare classes?

Accuracy ignores imbalance. If the "churn" class appears in 2% of cases, the model can predict "all good" and get 98% accuracy — but the business loses clients. Solution: F1 macro (averaged over all classes) or weighted F1. For NER — strict entity F1 (exact matches only). We guarantee: after choosing the correct metric, model quality becomes measurable and predictable.

Text Classification: From BERT to Distillation

BERT-like models are the standard for classification. ruBERT-base or ruBERT-large from DeepPavlov for Russian. multilingual-e5-large — for multiple languages in one pipeline. XLM-RoBERTa-large — a strong multilingual backbone.

Fine-tuning for classification: add a classification head on top of the [CLS] token, train for 3-5 epochs with lr=2e-5, weight decay=0.01. For imbalance — weighted CrossEntropyLoss or focal loss with gamma=2.0. Contact us — we will show a code snippet.

Imbalance case study. Dataset — 3,000 examples, imbalance 1:20. Solution: class_weight via sklearn + CrossEntropyLoss. Additionally — augmentation of rare classes via backtranslation (ru→en→ru through MarianMT). Recall for rare classes rose from 0.31 to 0.67 with a slight drop in accuracy (76%→74%). Full NLP development end-to-end took 3 weeks.

Distillation for production. BERT-large gives F1 0.89, but inference on CPU — 180ms. Distillation into DistilBERT or ruBERT-tiny2 reduces latency to 25ms with F1 0.84. Export to ONNX Runtime provides an additional 1.5-2x speedup. DistilBERT achieves 7x lower latency than BERT-large with only a 5% drop in macro F1 – a typical production trade-off.

How to choose between BERT and LLM for your task?

For most classification and extraction tasks, BERT-sized models offer the best trade-off between cost and performance. Shift to LLMs only when the task demands generation, complex reasoning, or zero-shot generalization.

NER: Named Entity Recognition

NER — extracting persons, organizations, locations, dates, amounts, document numbers. For general categories (PER, ORG, LOC), pre-trained models work well. For specialized ones (medical terms, legal concepts) — fine-tuning is needed.

Data annotation. The main cost of an NER project. For a quality model — 500-2,000 labeled sentences per entity type. Tools: Label Studio (open source) or Prodigy (by spaCy creators). IOB2 format — standard.

Architecture. Token classification on top of BERT: each token gets a label (B-PER, I-PER, O). spaCy 3.x with transformer pipeline — a convenient production choice.

Nested entities. Standard IOB models cannot handle nested entities (organization inside an address). For such tasks — span-based NER: SpanBERT or SpERT. More complex but correct.

Post-processing is mandatory. The model predicts tokens — normalized entities are needed. Date — dateparser. Amounts — regex + validation. Names — deduplication via rapidfuzz. Included in our standard delivery.

Sentiment Analysis and Opinion Mining

Binary classification positive/negative works out of the box with BERT. Complexity — aspect-based sentiment analysis (ABSA): "the restaurant has good food but terrible service." For ABSA: aspect extraction (NER) + sentiment per aspect. Joint models BERT-for-ABSA — quality on Russian data is lower due to dataset scarcity. RuSentiment, SentiRuEval — main resources.

For production with simple positive/negative/neutral: distil models are enough. Three classes, balanced dataset, 2,000+ examples — F1 macro 0.82-0.87 in 1-2 days.

Text Summarization

Extractive summarization (select sentences) — TextRank or BM25 without training. Fast, no hallucinations. Good for long documents.

Abstractive (generates new text) — seq2seq: mT5, mBART, FRED-T5, ruT5-large. For production via LLM API (GPT-4, Claude) — often the best cost/quality/speed trade-off.

Embeddings: Vector Representations of Text

Embeddings are the foundation of semantic search, deduplication, clustering, RAG. Quality critically affects downstream tasks.

Models. E5-large-v2, BGE-M3, multilingual-e5-large — strong multilingual embedders. sentence-transformers/paraphrase-multilingual-mpnet-base-v2 — fast option. For Russian: ru-en-RoSBERTa (Skoltech) performs well on semantic textual similarity.

Embedding quality evaluation uses the MTEB benchmark as standard. But top results on MTEB don't guarantee success on a domain dataset — we build domain-specific eval.

Fine-tuning embeddings. If standard models don't give the required Recall@k — contrastive learning on domain pairs with MultipleNegativesRankingLoss. How to perform this for domain data:

1. Collect 500–2,000 semantically similar pairs from your domain.
2. Apply MultipleNegativesRankingLoss with a batch size of 32–64.
3. Train for 1–3 epochs using AdamW (lr=2e-5).
4. Evaluate Recall@k on a held-out domain test set.

This approach yields a 5–15% improvement in Recall@k in practice.

Dimensionality and storage. E5-large: 1024 dim, float32 — 4KB per vector. For 10M documents — 40GB. Quantization int8 reduces to 10GB. FAISS IVF_PQ — more compact but with losses. Included in our deployment recommendations.

Information Extraction

Structured extraction is a frequent task. Examples: key contract terms, technical characteristics, dates and amounts from invoices.

1. Regex + rule-based. For INN, OGRN, amounts, dates — more reliable than neural networks. No data required.
2. NER + post-processing. For variable formats.
3. LLM with structured output. GPT‑4 / Claude with JSON schema — for complex documents. Cost: minimal per document. For 10k+ documents/day — we calculate the economics.

We guarantee a hybrid: regex/NER for typical fields + LLM for edge cases. Our guarantee is backed by years of production experience and more than 30 projects.

Work Stages

Prototype on existing data — 1-3 weeks. Production system with CI/CD — 1.5–2.5 months. Cost is calculated individually — get a consultation for a project estimate.

What's Included

• Model and pipeline architecture documentation
• Access to the model via REST API (FastAPI + ONNX)
• Client team training (2-hour webinar + Q&A)
• Accuracy guarantee on the agreed test set
• Months of post-delivery support (bug fixes, adaptation to new data)

Our Experience

Years of NLP projects from classification to RAG systems. The team includes ML engineers experienced with Hugging Face, spaCy, LangChain, MLOps. We use vLLM, Kubeflow, Weights & Biases — a production stack, not toys. Contact us to evaluate your NLP project within two days — request a free consultation on your text processing pipeline.