Extractive QA concept in NLP - Model Pipeline Trace

Model Pipeline - Extractive QA concept

This pipeline finds answers by picking exact text spans from a given passage based on a question. It reads the passage and question, then highlights the answer inside the passage.

Data Flow - 5 Stages

1Input Data

1000 samples x 2 texts (question, passage)→Receive pairs of question and passage texts→1000 samples x 2 texts (question, passage)

Question: 'Where is the Eiffel Tower located?'; Passage: 'The Eiffel Tower is in Paris, France, and is a famous landmark.'

↓

2Tokenization

1000 samples x 2 texts→Split texts into tokens (words or subwords)→1000 samples x 2 token lists (question tokens, passage tokens)

Question tokens: ['Where', 'is', 'the', 'Eiffel', 'Tower', 'located', '?']; Passage tokens: ['The', 'Eiffel', 'Tower', 'is', 'in', 'Paris', ',', 'France', ',', 'and', 'is', 'a', 'famous', 'landmark', '.']

↓

3Input Encoding

1000 samples x 2 token lists→Convert tokens to numerical vectors using embeddings→1000 samples x 2 sequences of vectors (e.g., 768-dim)

Question vectors: [[0.1, 0.3, ...], ...]; Passage vectors: [[0.2, 0.4, ...], ...]

↓

4Model Forward Pass

1000 samples x 2 sequences of vectors→Use a neural network (e.g., BERT) to predict start and end positions of answer in passage→1000 samples x 2 probability distributions over passage tokens (start and end)

Start probs: [0.01, 0.02, 0.7, 0.1, ...]; End probs: [0.01, 0.02, 0.1, 0.6, ...]

↓

5Answer Extraction

1000 samples x 2 probability distributions→Select token span with highest combined start and end probabilities→1000 samples x 1 text span (answer)

Answer: 'Paris, France'

Training Trace - Epoch by Epoch

Loss: 1.2 |****
      0.8 |******
      0.5 |*********
      0.35|***********
      0.3 |************
       Epochs -> 1 2 3 4 5

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Model starts learning, loss is high, accuracy low
2	0.8	0.60	Loss decreases, accuracy improves as model learns to locate answers
3	0.5	0.75	Model shows good understanding, loss continues to drop
4	0.35	0.82	Training converges, accuracy stabilizes near 82%
5	0.30	0.85	Final epoch, model achieves strong performance

Prediction Trace - 5 Layers

Layer 1: Tokenization

Layer 2: Input Encoding

Layer 3: Model Forward Pass

Layer 4: Answer Extraction

Layer 5: Detokenization

Model Quiz - 3 Questions

What does the model predict to find the answer in the passage?

A Only the first word of the passage

B Start and end positions of the answer span

C The entire passage as the answer

D A summary of the passage

Key Insight

Extractive QA models learn to locate exact answer spans by predicting start and end positions in the passage. As training progresses, the model improves by reducing loss and increasing accuracy, enabling precise answer extraction from text.