Computational Linguistics and Probability Theory Notes

Computational Linguistics

• Definition: A field that combines linguistics and computer science, focusing on statistical description of natural language.
• Key Concepts: Involves probability and information theory as foundational elements to evaluate language processing.

Random Experiments and Events

• Random Experiment: An experiment whose outcomes cannot be precisely determined beforehand; can be repeated multiple times.
  • Example: Tossing a die.
• Random Phenomenon: Statement about the outcome of a random experiment, determined post-experiment.
  • Examples of Outcomes: Even number, number greater than 3.
• Elementary Phenomenon: Base outcomes that cannot be further decomposed.
  • Sample Space: All elementary phenomena $ ext{Ω}$.

Probability Basics

• Probability of Event A: Given by $P(A) = \frac{k}{N}$ where:
  • $N$: Total number of trials.
  • $k$: Number of occurrences of event A.
• Characteristics:
  • Probability values range: $0 ext{ ≤ } P(A) ext{ ≤ } 1$.
  • Certain events have probability = 1 $P(Ω) = 1$.
• Sum of Probabilities for Incompatible Events: $P(A \v B) = P(A) + P(B)$.
• Determining Probabilities: Can utilize combinatorial methods or statistical measurements from real data.

Joint and Conditional Probability

• Joint Probability: Denoted as $P(A ∩ B) = P(A, B)$.
• Conditional Probability: Denoted as $P(A | B) = \frac{P(A ∩ B)}{P(B)}$ which estimates from relative frequencies.
  • Rearranged as: $P(A|B) = \frac{c(A,B)/N}{c(B)/N} = \frac{c(A,B)}{c(B)}$.

Word Probability Examples

• Examining Word Probabilities:
  • For the word "today" in Czech, labelled as event A.
  • Event B as occurrence of the word "respektive".
• Calculating Probabilities:
  • $P(A) = \frac{c(A)}{N_w}=\frac{2}{4}=\frac{1}{2}$.
  • $P(B) = \frac{c(B)}{N_w}=\frac{1}{4}$.
• Pair Probabilities:
  • $P(A,B) = \frac{c(A,B)}{N_{w-1,w}} = \frac{1}{4}$.
  • Conditional Probability: $P(B|A) = \frac{1/4}{1/2} = \frac{1}{2}$.

Random Variables

• Definition: Variable whose value is determined by the outcome of a random experiment, with discrete or continuous values.
  • Discrete Random Variable: Countable outcomes.
  • Continuous Random Variable: Unbounded possible values.
• Probability Distribution: Function describing the probabilities of individual outcomes. The distribution function $F(x) = P(X ≤ x)$.

Entropy and Information

• Concept of Entropy: Measures uncertainty in the outcome of a random variable.
  • More information leads to lower entropy.
  • Formula: For a coin toss, $H(X) = - \big( P( ext{head}) imes ext{log}<em>2(P( ext{head})) + P( ext{tail}) imes ext{log}</em>2(P( ext{tail})) \big)$.
  • Example with a 32-sided die:
  • Calculation shows entropy is $H(X) = 5$ bits (larger uncertainty).
• Entropy Properties:
  • Entropy (H) is non-negative, $0 ≤ H(X)$.
  • Uniform distribution has maximum entropy

Perplexity

• Definition: Measure of how difficult it is to predict outcomes of random variables.
• Examples of Perplexity:
  • Variable with 2 outcomes: $H(X)= 1 ext{ bit}, G(X) = 2$.
  • Variable with 32 outcomes of varying probabilities has different perceptual difficulties in prediction.

Joint and Conditional Entropy

• Joint Entropy: Measures total uncertainty of two random variables $X$ and $Y$.
• Conditional Entropy: Measures remaining uncertainty of one variable given knowledge of the other, calculated through $H(Y|X)$.

General Properties of Language Entropy

• English Example: Different entropies depending on character prediction (
  o character knowledge, previous, pairs, etc.).
• Entropy Decrease: Observing previous events lowers entropy.

Kullback-Leibler Divergence** (Relative Entropy)

• Measures the difference between two probability distributions $ ext{D}(p ext{||} q)$.
• Not symmetric: $D(p ext{||} q) <br />\neq D(q ext{||} p)$.

Cross Entropy

• Indicates bits required to describe one probability distribution using another. $C(p, q) = - ext{∑}_x p(x) ext{log}(q(x))$.

Mutual Information

• Quantifies the reduction in uncertainty of one random variable given knowledge of another. $I(X, Y) = H(X) - H(X|Y)$. If independent: $I(X, Y) = 0$.