Overview
Teaching computers to speak human. From the early days of rule-based systems to the modern era of Deep Learning and LLMs (Large Language Models).
Core Idea
Statistical vs. Symbolic:
- Symbolic (Old School): Hand-coding grammar rules. (If Noun + Verb…). Good for precision, bad for ambiguity.
- Statistical (Modern): Feeding the computer billions of words and letting it learn the patterns (probabilities). “The cat sat on the…” (Mat: 90%, Hat: 5%).
Formal Definition (if applicable)
Word Embeddings (Vectors): Representing words as numbers in a multi-dimensional space. “King” - “Man” + “Woman” = “Queen”. This allows computers to understand meaning and analogy.
Intuition
- Machine Translation: Google Translate. Used to be word-for-word (terrible). Now uses Neural Networks to translate whole sentences (fluent).
- Speech Recognition: Siri/Alexa. Turning sound waves into text.
Examples
- LLMs (GPT, Gemini): Predicting the next word. It turns out if you do this well enough, you get reasoning, coding, and poetry.
- Sentiment Analysis: Reading tweets to see if people are happy or angry about a movie.
Common Misconceptions
- “The AI understands.” (It manipulates symbols/probabilities. Whether it has “understanding” is a philosophical debate).
- “It’s solved.” (Still struggles with sarcasm, deep context, and hallucination).
Related Concepts
- Turing Test: Can a machine fool a human?
- Corpus Linguistics: Analyzing huge databases of text to see how language is actually used.
Applications
- Search: Google.
- Accessibility: Screen readers for the blind.
- Customer Service: Chatbots.
Criticism / Limitations
Bias. If the training data is biased (sexist/racist), the AI will be too.
Further Reading
- Jurafsky & Martin, Speech and Language Processing
- Manning & Schütze, Foundations of Statistical Natural Language Processing