Page 1 – The Transformer Architecture

1.1 Origin of Transformers

The Transformer model was introduced by Vaswani et al. (2017) in “Attention is All You Need.” It eliminated recurrence entirely in favour of:

• Self-attention
• Parallel computation
• Improved long-range dependency modelling

1.2 Encoder–Decoder Design

The Transformer contains two main components:

• Encoder – maps input sequence to contextual representations.
• Decoder – autoregressive generator for output sequences.

1.3 Multi-Head Attention

Multi-head attention allows the model to learn multiple relationships simultaneously.

Q = XW_Q
K = XW_K
V = XW_V
Attention = softmax(QKᵀ / √dₖ) V
    

1.4 Positional Encoding

Because Transformers have no inherent recurrence, they inject order using positional encodings:

• Sinusoidal encodings
• Learned positional embeddings
• Rotary and ALiBi encodings in modern models

1.5 Feed-Forward Networks

Each Transformer layer includes a position-wise feed-forward network:

FFN(x) = max(0, xW₁ + b₁)W₂ + b₂

1.6 Layer Normalisation & Residual Connections

Residual connections + layer normalisation stabilise deep architecture training.

1.7 Why Transformers Replaced RNNs

• Better global context modelling
• Parallelisation → faster training
• Higher performance on seq2seq tasks
• Scalable to billions of parameters

Page 2 – Self-Supervised Pretraining & Autoregressive Models

2.1 The Self-Supervised Paradigm

Instead of requiring labelled data, LLMs are trained on self-supervised objectives:

• Predict masked tokens (BERT)
• Predict next token (GPT)
• Reconstruct original text from corrupted input (BART, T5)

2.2 Masked Language Modelling (MLM)

MLM masks 15% of tokens and asks the model to fill in blanks. Effective for bidirectional understanding.

2.3 Autoregressive Language Modelling

GPT-style models learn:

P(wₜ | w₁, …, wₜ₋₁)

This enables open-ended text generation and reasoning.

2.4 Encoder-Only vs. Decoder-Only vs. Encoder–Decoder

• Encoder-only (BERT) → classification + extraction tasks
• Decoder-only (GPT) → generative tasks
• Encoder–decoder (T5, BART) → translation, summarisation

2.5 Scaling Laws

Research shows predictable improvement as we scale:

• Model size (parameters)
• Dataset size
• Training compute

2.6 Emergent Behaviour

Large-scale models display properties not present in small models:

• Chain-of-thought reasoning
• Code generation
• Tool use
• In-context learning

Page 3 – Large Language Models (LLMs)

3.1 What Is an LLM?

A Large Language Model is a Transformer-based model with billions of parameters, trained on massive corpora to learn general language representations and reasoning.

3.2 Key Capabilities

• Natural language understanding
• Text generation
• Translation & summarisation
• Information retrieval
• Reasoning & chain-of-thought
• Coding & mathematical problem solving

3.3 Instruction Tuning

Models like ChatGPT, LLaMA-2, PaLM, and Mistral use:

• Instruction tuning – training on high-quality instruction–response pairs.
• Supervised fine-tuning (SFT)

3.4 Reinforcement Learning from Human Feedback (RLHF)

RLHF aligns model outputs with human preferences:

1. Generate model responses.
2. Rank by human evaluators.
3. Train a reward model.
4. Optimise using PPO or similar algorithms.

3.5 Retrieval-Augmented Generation (RAG)

RAG systems enhance LLMs with external knowledge:

• Vector databases
• Document retrieval
• Real-time information access

3.6 Limitations of LLMs

• Hallucination (fabricated content)
• Bias & stereotypical associations
• Lack of interpretability
• High computational cost
• Inconsistent reasoning in edge cases

Page 4 – Prompt Engineering

4.1 Why Prompts Matter

LLMs behave differently depending on how the input is phrased. Prompt engineering improves control, reliability, and output quality.

4.2 Types of Prompts

• Zero-shot prompts – no examples provided.
• Few-shot prompts – include demonstrations.
• Chain-of-thought prompts – encourage step-by-step reasoning.
• Role-based prompts – instruct model to adopt a persona.

4.3 Example of Few-Shot Prompt

Translate to Thai:
Input: "I am hungry"
Output: "ฉันหิว"
    

4.4 Prompt Patterns

• Instruction → Input → Output
• Role → Task → Constraints
• Context → Question

4.5 Avoiding Prompt Pitfalls

• Avoid ambiguous instructions.
• Specify output format.
• Include examples for complex tasks.
• Use delimiters to separate content.

4.6 Advanced Prompting

• Self-consistency – majority reasoning.
• Generated knowledge prompting – ask model to generate background info first.
• Tool-use prompting – combine with external APIs.

Page 5 – Applications, Risks, Deployment & Future Directions

5.1 Real-World Applications

• Healthcare NLP (clinical summarisation, coding)
• Legal text analysis
• Customer support automation
• Search engines & conversational interfaces
• Education (tutoring, feedback generation)
• Software development assistants
• Scientific literature analysis

5.2 Risks & Safety Concerns

Deploying LLMs introduces critical risks:

• Hallucinations – confident but incorrect answers
• Bias & fairness issues – inherited from training data
• Privacy leakage – memorisation of sensitive training examples
• Security vulnerabilities – prompt injection, jailbreaking
• Misinformation scaling

5.3 Responsible Deployment

• Model monitoring
• Human-in-the-loop systems
• Evaluation on safety benchmarks
• RAG-based verification
• Transparent documentation

5.4 Future Directions

• Multimodal LLMs (text, image, audio, video)
• Agentic AI – tool use, reasoning with external environments
• Continual learning and memory-based architectures
• Neural–symbolic integration for reasoning
• On-device LLMs for privacy-first applications

5.5 Summary

Module 5 concludes the NLP course by presenting the architectures and methodologies that define modern NLP—Transformers, LLMs, prompt engineering, and real-world deployment considerations. These models dominate today’s AI landscape and continue to reshape the future of language technologies.