Intent Classification Bot

NLP chatbot with bag-of-words and neural network classification.

PythonNLTKTensorFlowBoWNeural Network

Demo Conversation

Intent Detection

weather94% conf

"What's the weather like?"

"Tell me the forecast"

"Is it going to rain?"

goodbye97% conf

"Bye"

"See you later"

"Thanks, bye!"

NLP Pipeline

Tokenization

Split sentence into words

"What's the weather?" → ["what", "the", "weather"]

Stemming

Reduce to root forms (Porter Stemmer)

["running", "runs"] → ["run", "run"]

Bag-of-Words

Binary vector encoding

vocab=[hello,weather,bye] → [0,1,0]

Classification

Neural network predicts intent

[0,1,0] → weather (94% confidence)

Neural Network

Input

BoW Vector

Hidden 1

128 neurons

Hidden 2

64 neurons

Output

Softmax

Modular Intent Corpus

intents.json

{
  "intents": [
    {
      "tag": "greeting",
      "patterns": ["Hi", "Hello", "Hey"],
      "responses": ["Hello!", "Hi there!"]
    },
    {
      "tag": "weather",
      "patterns": ["Weather?", "Forecast"],
      "responses": ["Which city?"]
    }
  ]
}

Easy Updates

Add intents by editing JSON — no code changes needed

Fast Retraining

<1 minute training on small datasets

Confidence Threshold

Falls back to "I don't understand" when uncertain (<70%)

Design Decisions

BoW + NN over Transformers

Sub-50ms inference on CPU. Perfect for domain-specific intents where BERT/GPT would be overkill.

Stemming > Lemmatization

Faster, no POS tagging needed. Acceptable tradeoff for intent matching.

Dropout for Regularization

50% dropout prevents overfitting on small training datasets.

Softmax with Threshold

Better to say "I don't know" than confidently return wrong answers.

<50ms

Inference Time

0.7

Confidence Threshold

128×64

Hidden Layers

JSON

Intent Format