In the competitive landscape of 2026, response speed is no longer the sole determining factor in the credit and real estate sectors. The real challenge lies in precision and the ability to filter out noise. **Real estate lead qualification** has shifted from being a manual task performed by call centers to an automated process driven by Natural Language Processing (NLP) algorithms. In this technical guide, we will explore how to build a customized Named Entity Recognition (NER) system to extract structured data from unstructured conversations and integrate them directly into the BOMA CRM.

Why Entity Extraction Changes Real Estate Lead Qualification

Static forms on websites (Name, Surname, Phone) have increasingly lower conversion rates. Users prefer interacting via natural chats or voice messages. However, this generates unstructured data that is difficult to process. This is where **Semantic Entity Extraction** comes into play.

The goal is not just to understand the intent (e.g., “I want a mortgage”), but to extract specific slots of information necessary for calculating credit ratings or purchase feasibility. A well-designed system must identify:

• ENT_AMOUNT: The requested amount (e.g., “I need 200k”).
• ENT_LTV: The implied Loan-to-Value or property value.
• ENT_JOB_TYPE: The contract type (e.g., “permanent”, “flat-rate freelancer”).
• ENT_PROPERTY: Property type and energy class.

Prerequisites and Tech Stack

To follow this guide, intermediate knowledge of Python and Machine Learning principles is required. We will use the following stack, standardized for 2026:

• Language: Python 3.12+
• NLP Framework: Hugging Face Transformers, spaCy 4.x
• Base Models: UmBERTo (for Italian) or quantized versions of Llama-3-8B-Instruct for generative tasks.
• Backend: FastAPI for model exposure.
• Target CRM: BOMA (via REST API/Webhook).

Phase 1: Designing the Entity Schema

Before writing code, we must define what our model needs to look for. In the context of mortgages, the jargon is specific. A generic model would fail to distinguish between “down payment” and “installment”.

Let’s define the labels for our training dataset:

NER_TAGS = [
    "O",              # Outside (no entity)
    "B-REQ_AMOUNT",   # Start requested amount
    "I-REQ_AMOUNT",   # Inside requested amount
    "B-JOB_STATUS",   # Start job status
    "I-JOB_STATUS",   # Inside job status
    "B-PROPERTY_VAL", # Property value
    "B-INTENT_TIME"   # Desired timing (e.g., "closing by March")
]

Phase 2: Dataset Preparation and Fine-Tuning

To achieve precise real estate lead qualification, we cannot rely on generalist zero-shot models for massive extraction, as they are expensive and slow. The best solution is fine-tuning a BERT-based model.

1. Creating the Synthetic Dataset

If you do not have GDPR-compliant chat histories, you can generate a synthetic dataset using an LLM (like Meta AI Llama 3) to create thousands of variations of typical phrases:

“I am a state employee looking for a mortgage for a house worth 250,000 euros, I have 50k down payment.”

Annotate these phrases in the standard JSONL format for training (BIO format).

2. Fine-Tuning with Hugging Face

We will use dbmdz/bert-base-italian-xxl-cased as a base, being one of the highest-performing models on Italian syntax. Here is a simplified snippet for the training loop:

from transformers import AutoTokenizer, AutoModelForTokenClassification, TrainingArguments, Trainer

model_name = "dbmdz/bert-base-italian-xxl-cased"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name, num_labels=len(NER_TAGS))

args = TrainingArguments(
    output_dir="./boma-ner-v1",
    evaluation_strategy="epoch",
    learning_rate=2e-5,
    per_device_train_batch_size=16,
    num_train_epochs=5,
    weight_decay=0.01,
)

# Assuming 'tokenized_datasets' is already prepared
trainer = Trainer(
    model=model,
    args=args,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["validation"],
    tokenizer=tokenizer,
)

trainer.train()

This process adapts the model weights to specifically recognize terms like “refinancing”, “fixed rate”, or “consultant” in the context of the sentence.

Phase 3: Post-Processing and Normalization

The NER model returns tokens and labels. For real estate lead qualification, we need to transform "two hundred thousand euros" into 200000 (Integer). This normalization phase is critical for populating the database.

We use libraries like word2number or custom regex to clean the model output before sending it to the CRM.

Phase 4: Integration into BOMA CRM

Once the model is exposed via API (e.g., on a Docker container), we need to connect it to the lead inflow. Integration with BOMA usually happens via webhooks that trigger upon receipt of a new message.

Scoring and Routing Logic

Not all leads are equal. Using the extracted data, we can calculate a Lead Quality Score (LQS) in real-time:

• Lead A (Score 90/100): Complete data (Job, Amount, Property), LTV Immediate routing to Senior Consultant.
• Lead B (Score 40/100): Partial data, LTV > 95%, Fixed-term Contract. -> Routing to automatic Nurturing Bot.

Here is an example of a JSON payload to send to the BOMA APIs:

{
  "lead_source": "Whatsapp_Business",
  "message_body": "Hi, I would like info for a first home mortgage, I am a nurse",
  "extracted_data": {
    "job_type": "nurse",
    "job_category": "public_sector",
    "intent": "first_home_purchase"
  },
  "ai_score": 75,
  "routing_action": "assign_to_human"
}

Troubleshooting: Managing Hallucinations and Ambiguity

Even the best models can make mistakes. Here is how to mitigate risks:

1. Confidence Threshold: If the model extracts an entity with confidence lower than 85%, the system must mark the field as “To be verified” in the BOMA CRM, requiring human intervention.
2. Human-in-the-loop: Implement a feedback mechanism where real estate agents can correct labeling in the CRM. These corrected data must go back into the training dataset for monthly model re-training.
3. Dialect Management: Train the model on datasets that include regional colloquial expressions often used in informal chats.

Conclusions

Implementing an Entity Extraction system for real estate lead qualification is no longer an academic exercise, but an operational necessity. By automating the extraction of critical data (LTV, job, budget) and integrating them directly into BOMA, agencies can reduce first contact time from hours to seconds, assigning the most complex cases to the best consultants and leaving initial screening to AI.

Frequently Asked Questions