Implementing Text Summarization: Extractive and Abstractive Approaches

Which Approach Should You Choose for Text Summarization?

When summarizing a news feed with a volume of 10,000 articles per day, developers face a dilemma: extraction does not produce coherent text, while abstraction risks hallucinations. The solution is to combine methods and tune them to your data.

Key Challenges and Solutions

A major challenge is hallucinations in abstractive models. In legal documents, the cost of an error can reach millions, so we prefer a hybrid: extraction for safe fact extraction combined with fine-tuned T5 for the final text. Another issue is the limited context window. Documents exceeding 128k tokens must be chunked with 10% overlap and hierarchical summarization applied: first by sections, then a summary. Performance is also critical: under a load of 2000 req/s, latency p99 should not exceed 500 ms. We achieve this through vLLM, ONNX Runtime, and INT8 quantization. We also incorporate RAG for grounding summaries in external knowledge.

Our Approach: Stack and Case Study

In one project for a corporate news aggregator, we configured a hybrid system. The first pass was extractive sentence selection via TextRank with a similarity threshold of 0.7. The second was generating a summary from the selected sentences using IlyaGusev/rut5-base-absum, fine-tuned with LoRA on the company's news. Inference was accelerated via ONNX Runtime with dynamic quantization — FLOPS rose to 150 TFLOPS on a single A100. Result: ROUGE-1 0.52, BERTScore 0.68, latency p99 at 320 ms. This reduced reading time by 60%, saving the client approximately $15,000 annually in analyst hours.

Reasons for Errors in Abstractive Summarization

Models like GPT-4o are prone to hallucination when context is vague or contradictory. To minimize this, we use few-shot prompts with examples from your domain and add a constraint: "Do not use facts not present in the text." For critical scenarios, chain-of-thought is effective — the model first extracts facts and then formulates a summary. It is also important to consider that summarization quality directly depends on the representativeness of training data: fine-tuning on 10,000 documents from your domain can improve BERTScore by 0.05. In our tests, a hybrid scheme with a fine-tuned T5 outperforms pure extraction on ROUGE-1 by a factor of 1.3.

Choosing Between Extractive and Abstractive Summarization

Comparison of Approaches by Criteria

Process

1. Requirements analysis: content type, token volume, success metrics (ROUGE, BERTScore, latency).
2. Architecture selection: extractive, abstractive, or hybrid.
3. Training/fine-tuning: LoRA fine-tuning for abstractive model.
4. Integration: REST/gRPC API based on Docker image.
5. Testing: ROUGE/BERTScore + A/B test on live data.
6. Deployment: vLLM, ONNX, monitoring via Weights & Biases.

Estimated Timelines and Cost

Basic extractive summarization is set up in 3 working days and starts from $2,000. Abstractive with fine-tuning and production-ready integration — from 2 to 4 weeks, starting at $5,000. The cost is calculated individually: contact us — we will assess your project.

What's Included

• Architectural documentation (model card, pipeline diagram).
• Model code in a Docker image with ONNX runtime.
• Integration test (sample API call).
• Access to MLflow monitoring.
• MLOps pipeline setup.
• Team training (2-hour webinar with case studies).
• 3-month support guarantee.

Optimizing Latency for Production

An optimized ONNX model runs 2-3 times faster than base PyTorch. For high loads, we use vLLM with continuous batching — this reduces latency p99 to 300 ms at 1000 req/s. Order a pilot project — we will conduct an A/B test on your data.

Our team has over 5 years of experience, completed 5+ summarization projects, and served 10+ enterprise clients. We have been on the market for over 5 years and guarantee results that meet your metrics. Get a consultation — write to us.

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.