Text Analytics for Chat Dialogs: Analysis with SLA Guarantee

Text Analytics for Chat Dialogs: Analysis with SLA Guarantee

Operator review of dialogs consumes up to 40% of QA team time — while hidden trends remain unnoticed. We build a Text Analytics system that automatically turns hundreds of thousands of chats into structured business intelligence: not just statistics of contacts, but deep understanding — what customers are talking about, how topics change, where systemic problems arise.

Our experience: 5+ years of NLP solution development, 30+ projects for retail, fintech, and telecom. We guarantee SLA of 99.9% on processing time and classification accuracy of at least 90%. We process up to 5000 dialogs per hour on a single GPU — sufficient for Enterprise streaming.

Why Standard Reports Don't Give the Full Picture?

Traditional dashboards show the number of contacts but do not reveal context. A sharp increase in dialog count may be caused by an outage that is not visible in first-level metrics. Our system detects anomalies and links them to changes in topic, tone, and resolution — allowing response before the problem becomes widespread. On one project, we identified a 12% increase in product negativity three days before operators noticed it.

Compare two approaches: simple keyword-based frequency classification gives 50-60% accuracy and misses up to 40% of hidden patterns. Fine-tuned ML models (ruBERT, SDG) yield 80-94% F1 but require labeled data. LLM (GPT-4) achieves 95%+ without labels, but token cost is 5-10 times higher for mass processing. Our solution is hybrid: ML for routine classification (70% of volume) and LLM for complex cases (30%) — this gives the optimal price/quality ratio.

Architecture of the Analytics System

[Dialogs from helpdesk/CRM/bots]
    → [Enrichment: topic, sentiment, NER, summary]
    → [Storage: ClickHouse / BigQuery]
    → [Aggregation and data marts]
    → [Dashboards: Superset / Metabase / Grafana]
    + [Ad-hoc analysis: Jupyter / Python API]

Stack: ClickHouse for storage, Celery for batches, GPU batching for embeddings. Models: ruBERT for topics, fine-tuned SDG for sentiment, NER based on SpaCy. Batch processing time for 10,000 dialogs — 2 hours on a single V100.

Comparison of Approaches to Dialog Analysis

The hybrid scheme yields 93% accuracy at $0.08 — 6 times cheaper than pure LLM with comparable quality.

Text Analytics Deployment Stages

How to Implement Text Analytics Without Stopping Current Processes?

We use a double-run approach: first, the system runs in parallel with existing processes, we validate results for two weeks, then switch main dashboards to new data. This avoids downtime and provides accuracy confirmation on real data. Pilot project on 10,000 dialogs — 5 days.

What's Included in the Work?

• Architectural documentation and pipeline description.
• Configured dashboards (3-5 data marts) for key metrics.
• Fine-tuned models with model card and accuracy report.
• API access for ad-hoc queries.
• Team training (2-3 workshops) and technical support for the first 2 months.

Dialog Enrichment

Each dialog goes through an enrichment pipeline:

@dataclass
class EnrichedDialog:
    dialog_id: str
    timestamp: datetime
    topic: str              # first-level topic
    subtopic: str           # detail
    sentiment_score: float  # -1 to 1
    sentiment_trend: str    # "improving" | "stable" | "worsening"
    resolution: bool        # was the problem solved
    escalated: bool
    key_entities: dict      # products, numbers, amounts
    summary: str            # brief description in 1-2 sentences
    emotion: str            # frustrated | satisfied | neutral | confused
    agent_id: str
    duration_seconds: int
    message_count: int

Batch processing: 10,000 dialogs overnight via Celery + GPU batching for embedding tasks. In real-time mode, Kafka stream processes a dialog in 2-5 seconds.

Analytics Data Marts

Topic Mart: aggregation by topic, subtopic, date, segment. Answers: what causes contacts, how it changes over time. Allows noticing seasonal peaks two weeks ahead.

Sentiment Mart: average score by segment, product, operator. Where customers are most/least satisfied. For example, the mart showed that NPS for the Premium product dropped by 8 points due to response delays over 10 minutes.

Problem Mart: dialogs with resolution=False + negative sentiment — source of systemic issues. Typical insight: 30% of such dialogs contain the word "document" — meaning document processing needs automation.

Operator Mart: performance metrics per agent for QA. Comparing sentiment before and after training shows an average improvement of 15%.

Anomalies and Alerts

Automatic anomaly detection in dialog stream:

• Sudden spike in topic contacts → likely incident
• Sharp deterioration in product sentiment → quality problem
• Increase in resolution=False share by type → procedure changed or knowledge base info missing

Alerts in Slack/Telegram: automatically when thresholds are exceeded. Incident confirmation time — less than 1 minute.

What Does Implementation Give?

• Reduce dialog analysis time by 80% (from 40% manual work to automation).
• Save QA budget up to 40% through automation.
• Cut incident response time from hours to minutes.
• Increase classification accuracy from 50-60% (keywords) to 93% (hybrid).

SLA 99.9% on processing time — guarantee of pipeline stability.

Order a pilot analysis: process 10,000 dialogs in 5 days and see first insights. Get a consultation — we'll explain how the system fits into your current stack.

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.