Automatic Action Item Extraction from Meeting Transcripts

A one-hour meeting, a 12-page transcript — and 20 minutes to extract tasks. A typical story: the recording mentions deadlines, assignees, but nothing ends up in Jira. Manual parsing of transcripts takes time and generates errors: missed tasks, incorrect deadlines. We solve this by automatically extracting Action Items with over 92% accuracy and reducing manual effort by 70%. Our clients save an average of 15,000–20,000 rubles per month by eliminating manual review. In one case, a client saved 50,000 rubles in the first month after deployment. Without a two-stage classification, models confuse discussions and tasks — for example, the phrase "We need to discuss the budget" is not a task but a topic. Our approach builds a robust pipeline: first classify fragments, then structure only tasks.

Improving Action Item Extraction Accuracy with Two-Step Approach

A direct prompt instructing "find all tasks" produces a lot of noise — the model includes discussions and questions as tasks. For instance, the phrase "We need to discuss the budget" is not an Action Item but a topic. The best approach is a dual-phase one:

1. Phrase Classification — the model labels transcript fragments as action_item, decision, question, discussion.
2. Structuring — only fragments of type action_item are processed to extract fields.

class ActionItem(BaseModel):
    task: str                    # task description
    assignee: str | None         # assignee name (if mentioned)
    deadline: str | None         # deadline (if mentioned)
    context: str                 # original quote from transcript
    confidence: float            # model confidence

Comparison with single-stage extraction:

Our two-step method is 1.5 times more accurate than single-stage extraction (92% vs 60%). It also reduces false positives by 4.4 times (35% to 8%), leading to 3.3 times less manual effort (from 70% to 20%).

Advantages of the Dual-Phase Approach over Direct Extraction

The two-step method allows separating actual tasks from hypothetical discussions. We use custom prompts with few-shot examples and chain-of-thought for classification. For assignee mapping, we use fuzzy matching based on embeddings. This ensures robustness to synonyms and name abbreviations. According to research, two-step classification improves accuracy by 30% compared to single-stage prompting.

Handling Uncertainty

Transcripts contain conditional obligations: "We should do", "Maybe Ivan will handle it". The model must distinguish:

• Clear obligation: "Peter, do it by Friday" → confidence 0.95
• Potential task: "We need to sort out this issue" → confidence 0.6, flagged for review

Action Items with confidence < 0.7 are placed in a separate Needs Clarification section.

What Quality Metrics Do We Guarantee?

During the testing phase, we conduct A/B comparison on your data. Target metrics: precision >90%, recall >85% after threshold tuning. For each project, we establish a baseline and achieve at least a 15% improvement over single-stage prompting. Implementation experience shows that the pipeline consistently delivers the stated accuracy.

Handling Low-Quality Transcripts

For noisy audio, we apply preprocessing: removing repetitions, normalizing noise, and segmenting utterances. If the overall certainty of a task is below 0.7, it is sent for manual review. For poor audio, we additionally use an ASR model (e.g., Whisper large-v3). This improves recognition accuracy and thereby extraction quality.

Configuring Tracker Integration

Automatic task creation in Jira / Linear / Asana / Trello via API after user confirmation (or automatically for tasks with confidence > 0.9). Assignee is mapped to real users via fuzzy matching by name. We also provide a webhook for custom integration.

Process and Timeline

1. Analysis — we study the structure of your meetings, typical phrases, and task formats.
2. Design — we choose the architecture (LLM, vector database, microservices).
3. Implementation — we write the classification and extraction pipeline, configure confidence thresholds.
4. Testing — A/B test on a sample, achieving precision >90%, recall >85%.
5. Deployment — launch in your infrastructure (ONNX Runtime to reduce latency).

Timeline: 5 to 10 business days for basic implementation. For complex cases, individually. Order a test run on your data — it's free. Get a consultation from an engineer to configure the solution for your infrastructure. Submit a request — and we'll demonstrate the result on your real transcripts. With over 5 years of experience in NLP and 50+ successful deployments, we guarantee quality.

What's Included

• Analysis of your transcripts and model tuning to your domain
• Service deployment (API or batch processing)
• Integration with task tracker
• Testing on historical data
• Documentation and team training
• Support for 2 weeks after launch

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.