effability

see also language

Transitions

1. Transition first from rule based learning to statistical learning
2. Rise of semantic parsing: statistical models of parsing
3. Then, moving from semantic parsing to large models—putting decision making and language modeling into the same bubble

Importance of LLMs

• They are simply better at understanding language inputs
• They can generate structured information (i.e. not just human language, JSONs, etc.)
• They can perform natural language “reasoning”—not just generate

(and natural language generation, abv)

What makes language modeling hard: resolving ambiguity is hard.

“the chef made her duck”

Contents

Basic Text Processing

• regex
• ELIZA
• tokenization and corpus
  • Herdan’s Law
• text normalization
  • tokenization + Subword Tokenization
    • BPE
  • Word Normalization
    • lemmatization through morphological parsing
    • only take stems from morphemes: porter stemmer
  • sentence segmentation
• N-Grams

Edit Distance

DP costs \(O(nm)\), backtrace costs \(O(n+m)\).

• minimum edit distance
  • weighted edit distance
• backtracing

Ngrams

• N-Grams
• Markov Assumption
  • Unigrams
  • Backoff and Stupid Backoff
  • Interpolation
  • OOV Words
• Model Evaluation
  • perplexity
  • open vocabulary

Text Classification

• Text Classification
  • Bag of Words
  • Naive Bayes
  • Naive Bayes for Text Classification
    • Binary Naive Bayes
  • Lexicon
• Naive Bays Language Modeling
• Harmonic Mean
  • Macroaverage and Microaverage

Logistic Regression

• Generative Classifier vs Discriminate Classifier
• Logistic Regression Text Classification
  • decision boundary
• cross entropy loss
• stochastic gradient descent

Information Retrial

• Information Retrival
  • Term-Document Matrix
  • Inverted Index + postings list
    • Boolean Retrieval
    • positional index

Ranked Information Retrial

• Ranked Information Retrieval
  • feast or famine problem
  • free text query
• score
  • Jaccard Coefficient
  • log-frequency weighting
  • document frequency ("idf weight")
  • TF-IDF
    • SMART notation
• vector-space model

Vector Semantics

• sense
  • principle of contrast
  • word relatedness
    • semantic field
  • synonymy and antonyms
  • affective meaning
• vector semantics
  • transposing a Term-Document Matrix
  • term-term matrix
  • word2vec
    • skip-gram with negative sampling

POS and NER

• POS Tagging
• NER Tagging

Dialogue Systems

• Dialogue
• Chatbot
  • PARRY

Recommender Systems

• Recommender System

Dora

• Dora

Neural Nets

• Neural Networks

The Web

• Web Graph
• Social Network

A machine learning model: input — last n words, output — probabilist distribution over the next word. An LM predicts this distribution (“what’s the distribution of next word given the previous words):

\begin{equation} W_{n} \sim P(\cdot | w^{(t-1)}, w^{(t-2)}, \dots, w^{(1)}) \end{equation}

By applying the chain rule, we can also think of the language model as assigning a probability to a sequence of words:

\begin{align} P(S) &= P(w^{(t)} | w^{(t-1)}, w^{(t-2)}, \dots, w^{(1)}) \cdot P(w^{(t-1)} | w^{(t-2)}, \dots, w^{(1)}) \dots \\ &= P(w^{(t)}, w^{(t-1)}, w^{(t-2)}, \dots, w^{(1)}) \end{align}

agents that uses the language to act on behave of another person or group.

Challenges

See Challenges of Language Model Agents

Methods

ReAct

See ReAct

Aguvis

Take the AgentNet dataset, and then tune a vison LM to roll out the rest of the sequence of actions given screenshots as input on top of a Qwen base model.

We can also add on top Chain of Thought to get more thinking as well.

Formulations

OSWorld

A unified task setup and evaluation.