Python Slot Filling and Extraction — Core Concepts

What Is Slot Filling?

Slot filling is the process of extracting specific pieces of information (entities) from user messages and mapping them to predefined fields (slots) that a chatbot needs to complete a task. It bridges the gap between free-form human language and the structured data that APIs and databases require.

How It Works

Step 1: Entity Extraction

The bot scans the user’s message and identifies spans of text that represent meaningful data. “Book a flight to Berlin on March 20 for 2 passengers” yields:

Text	Entity Type	Slot
Berlin	city	destination
March 20	date	travel_date
2	number	passenger_count

Entity extraction methods range from simple (regex, keyword lists) to complex (neural sequence labeling).

Step 2: Slot Mapping

Extracted entities are mapped to the bot’s form. Sometimes the mapping is direct — a city entity fills the destination slot. Other times it requires context: “I want to fly from Berlin to London” has two cities, and the bot must figure out which is the origin and which is the destination.

Step 3: Validation

Raw extracted values need validation. “February 30” is not a valid date. “Zero passengers” does not make sense. The bot validates each slot value and re-prompts if something is wrong.

Step 4: Prompting for Missing Slots

If the user’s message does not provide all required information, the bot asks follow-up questions. The dialog manager tracks which slots are filled and which are still empty, generating targeted prompts for the gaps.

Entity Extraction Approaches

Rule-Based

Regular expressions and lookup tables handle well-structured entities:

• Dates: Regex or dateutil parser for “March 20”, “next Friday”, “3/20/2026”
• Numbers: Regex for digits, word-to-number conversion for “two”, “three hundred”
• Enums: Lookup tables for known values like pizza sizes or airline codes

Rules are fast, predictable, and easy to debug — ideal for entities with limited, well-known formats.

Statistical (CRF)

Conditional Random Fields label each token in a sequence. They consider the surrounding tokens, so “fly to Berlin” correctly tags “Berlin” as a destination while “Berlin Wall history” might not. spaCy and sklearn-crfsuite both support CRF-based extraction in Python.

Neural (Transformer-Based)

Fine-tuned BERT or similar models treat entity extraction as token classification (BIO tagging). Each token gets a label: B-destination (beginning), I-destination (inside), or O (outside). This captures complex patterns but requires labeled training data.

Multi-Turn Slot Filling

Real conversations fill slots across multiple messages:

1. User: “I want to order a pizza”
2. Bot: “What size?” → fills intent=order_pizza
3. User: “Large” → fills size=large
4. Bot: “What toppings?” → size is filled, toppings is empty
5. User: “Pepperoni and mushrooms” → fills toppings

The dialog manager maintains a running frame and decides which slot to ask about next. The order of prompts can be fixed (always ask size first) or dynamic (ask for whatever the user is most likely to provide).

Slot Corrections and Overwrites

Users change their minds. “Actually, make it a medium” should overwrite the size slot from “large” to “medium.” A good slot-filling system detects correction intent and updates the appropriate slot without resetting the entire form.

Common Misconception

Slot filling is not just entity extraction. Extraction finds the data in text; slot filling is the full loop — extraction, mapping to the right field, validation, tracking across turns, and handling corrections. Many tutorials focus only on extraction and ignore the multi-turn state management that makes slot filling work in real products.

Python Libraries

• spaCy: Built-in NER for common entity types (people, places, dates, organizations). Custom models for domain-specific entities.
• Duckling (Facebook): A rule-based system that excels at dates, times, amounts, and measurements. Runs as a separate service with a Python wrapper.
• Rasa: Integrates entity extraction with slot filling in its dialog management. Slots are automatically filled from extracted entities.
• dateutil / parsedatetime: Python libraries for parsing human-readable dates into datetime objects.

The one thing to remember: Slot filling is the complete cycle of extracting details from text, placing them in the right form fields, validating them, and asking follow-up questions — turning a free-form conversation into structured, actionable data.