Why QA systems extract answers in NLP - Model Pipeline Impact

Model Pipeline - Why QA systems extract answers

This pipeline shows how a Question Answering (QA) system finds and extracts the exact answer from a given text passage. It starts with the question and passage, processes the text, trains a model to locate answers, and finally predicts the answer span.

Data Flow - 5 Stages

1Input Data

1000 samples (question + passage pairs)→Collect question and passage text pairs→1000 samples (question + passage pairs)

Question: 'What is the capital of France?' Passage: 'France's capital is Paris, known for the Eiffel Tower.'

↓

2Preprocessing

1000 samples (question + passage pairs)→Tokenize text into words and convert to numerical IDs→1000 samples x 2 sequences (question tokens, passage tokens)

Question tokens: ['What', 'is', 'the', 'capital', 'of', 'France', '?'] Passage tokens: ['France', "'s", 'capital', 'is', 'Paris', ',', 'known', 'for', 'the', 'Eiffel', 'Tower', '.']

↓

3Feature Engineering

1000 samples x 2 sequences→Create embeddings for tokens and position encodings→1000 samples x 2 sequences x 768 features

Embedding vector for 'Paris' token: [0.12, -0.05, ..., 0.33]

↓

4Model Training

1000 samples x 2 sequences x 768 features→Train QA model to predict start and end positions of answer in passage→Model with learned weights

Model learns to predict start=4, end=4 for answer 'Paris' in passage tokens

↓

5Prediction

New question + passage tokens→Model predicts answer span indices→Answer span indices (start, end)

Predicted answer span: start=4, end=4 corresponds to 'Paris'

Training Trace - Epoch by Epoch

Epochs
1 |***************
2 |************
3 |*********
4 |******
5 |****
Loss

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Model starts learning, loss high, accuracy low
2	0.9	0.60	Loss decreases, accuracy improves as model learns answer positions
3	0.7	0.72	Model better at locating answers, loss continues to drop
4	0.5	0.80	Good convergence, model accurately predicts answer spans
5	0.4	0.85	Training stabilizes with high accuracy and low loss

Prediction Trace - 4 Layers

Layer 1: Input tokenization

Layer 2: Embedding layer

Layer 3: Model prediction

Layer 4: Answer extraction

Model Quiz - 3 Questions

Why does the QA system tokenize the input text?

A To split text into manageable pieces for the model

B To remove stop words from the text

C To translate the text into another language

D To increase the length of the input

Key Insight

QA systems extract answers by learning to locate the exact position of the answer in a passage. This focused extraction helps provide precise and relevant answers rather than generating text from scratch.