Natural Language Processing

Author

Chao Ma

Published

June 6, 2026

Notes on natural language processing, from classical system-building approaches to modern neural and reinforcement-learning methods for language models.

CMU Advanced NLP Lecture 21: Mixture of Experts Sparse expert layers for language models: replacing dense FFNs, top-k routing, expert choice, hash routing, balance loss, Z-loss, and expert parallelism.

CMU Advanced NLP Lecture 16: Reinforcement Learning for LLMs RL for language models through one-step and token-level MDP views, policy gradient, group-based advantages, PPO, cold-start fine-tuning, verifiable rewards, math reasoning, and RLHF.

CMU Advanced NLP Lecture 9: Decoding Algorithms Greedy decoding, beam search, MAP pitfalls, sampling, top-k, top-p, temperature, and KV-cache speedups for language model generation.

CMU Advanced NLP Lecture 2: Learned Representation Tokenization, token embeddings, CBOW, Deep CBOW, training losses, and a small sentiment-classification experiment comparing shallow and deeper bag-of-words models.

CMU Advanced NLP Lecture 1: Introduction and Fundamentals NLP system-building methods from rules to supervised learning and reinforcement learning, plus the common ingredients of parameterization, learning, inference, softmax probabilities, and generation.

---
title: "Natural Language Processing"
author: "Chao Ma"
date: "2026-06-06"
---

Notes on natural language processing, from classical system-building approaches to modern neural and reinforcement-learning methods for language models.

---

::: {.content-grid}

::: {.content-card}
**[CMU Advanced NLP Lecture 21: Mixture of Experts](cmu-advanced-nlp-lecture-21-mixture-of-experts.qmd)**
Sparse expert layers for language models: replacing dense FFNs, top-k routing, expert choice, hash routing, balance loss, Z-loss, and expert parallelism.
:::

::: {.content-card}
**[CMU Advanced NLP Lecture 16: Reinforcement Learning for LLMs](cmu-advanced-nlp-lecture-16-reinforcement-learning-for-llms.qmd)**
RL for language models through one-step and token-level MDP views, policy gradient, group-based advantages, PPO, cold-start fine-tuning, verifiable rewards, math reasoning, and RLHF.
:::

::: {.content-card}
**[CMU Advanced NLP Lecture 9: Decoding Algorithms](cmu-advanced-nlp-lecture-9-decoding-algorithms.qmd)**
Greedy decoding, beam search, MAP pitfalls, sampling, top-k, top-p, temperature, and KV-cache speedups for language model generation.
:::

::: {.content-card}
**[CMU Advanced NLP Lecture 2: Learned Representation](cmu-advanced-nlp-lecture-2-learned-representation.qmd)**
Tokenization, token embeddings, CBOW, Deep CBOW, training losses, and a small sentiment-classification experiment comparing shallow and deeper bag-of-words models.
:::

::: {.content-card}
**[CMU Advanced NLP Lecture 1: Introduction and Fundamentals](cmu-advanced-nlp-lecture-1-introduction-fundamentals.qmd)**
NLP system-building methods from rules to supervised learning and reinforcement learning, plus the common ingredients of parameterization, learning, inference, softmax probabilities, and generation.
:::

:::