Corpora and Statistical Methods Albert Gatt Part 2 Probability distributions Example 1: Book publishing Case: publishing house considers whether to publish a new textbook on statistical NLP considerations include: production cost, expected sales, net profits (given cost) Problem: to publish or not to publish? depends on expected sales and profits if published, how many copies? depends on demand and cost Example 1: Demand & cost figures Suppose: book costs €35, of which: publisher gets €25 bookstore gets €6 author gets €4 To make a decision, publisher needs to estimate profits as a function of the probability of selling n books, for different values of n. profit = (€25 * n) – overall production cost Terminology Random variable In this example, the expected profit from selling n books is our random variable It takes on different values, depending on n We use uppercase (e.g. X) to denote the random variable Distribution The different values of X (denoted x) form a distribution. If each value x can be assigned a probability (the probability of making a given profit), then we can plot each value x against its likelihood. Definitions Random variable A variable whose numerical value is determined by chance. Formally, a function that returns a unique numerical value determined by the outcome of an uncertain situation. Can be discrete (our exclusive focus) or continuous Probability distribution For a discrete random variable X, the probability distribution p(x) gives the probabilities for each value x of X. The probabilities p(x) of all possible values of X sum to 1. The distribution tells us how much out of the overall probability space (the “probability mass”), each value of x takes up. Tabulated probability distribution No. copies sold Prod. cost Profits (X) Probability P(x) 5,000 £275,000 -£150,000 .20 10,000 £300,000 -£50,000 .40 20,000 £350,000 £150,000 .25 30,000 £400,000 £350,000 .10 40,000 £450,000 £550,000 .05 Plotting the distribution Uses of a probability distribution Computation of: mean: the expected value of X in the long run based on the specific values of X, and their probability NB: NOT interpreted as value in a sample of data, but expected (future) value based on sample. standard deviation & variance: the extent to which actual values of X will differ from the mean skewness: the extent to which our distribution is “balanced”, i.e. whether it’s symmetrical In graphics… Mean: expected value in the long run SD & variance: How much actual values deviate from mean overall Skewness: Symmetry or “tail” of our distribution Measures of expectation and variation The expected value (mean) The expected value of a discrete random variable X, denoted E[X] or μ, is a weighted average of the values of X weighted, because not all values x will have the same probability estimated by summing, for all values of X, the product of x and its probability p(x) E[ X ] xp( x) x More on expected value The mean or expected value tells us that, in the long run, we can expect X to have the value μ. E.g. in our example, our book publisher can expect longterm profits of: (-150,000 * .2) + (-50,000 * .4) + (150,000 * .25) + (350,000 * .1) + (550,000 * .05) = €50,000 Variance Mean is the expected value of X, E[X] Variance (σ2) reflects the extent to which the actual outcomes deviate from expectation (i.e. from E[X]) σ2 = E[(X – μ)2] = Σ(x – μ)2p(x) i.e. the weighted sum of deviations squared Reasons for squaring: eliminates the distinction between +ve and –ve makes it exponential: larger deviations are given more importance e.g. one deviation of 10 is as large as 4 deviations of 5 Standard deviation Variance gives overall dispersion or variation Standard deviation (σ) is the dispersion of possible outcomes; it indicates how spread out the distribution is. estimated as square root of variance 2 2 ( x ) p( x) x The book publishing example again Recall that for our new book on stat NLP, expected profit is £50,000 What’s the standard deviation? need to estimate (50000-x)2 for all x multiply by p(x) in each case take the square root of the result This is left as an exercise… Skewness The mean gives us the “centre” of a distribution. Standard deviation gives us dispersion. Skewness (denoted γ “gamma”) is a measure of the symmetry of the outcomes. (x ) x 3 3 Skewness, continued The formula calculates the average value of cubed deviations by the standard deviation cubed. Why cubed? The cube of a positive deviation is itself positive; that of a negative is itself negative. We want both, as we want to know deviations both to the left (-ve) and right (+ve) of the mean. Like the variance estimation, this emphasises large deviations in either direction (it’s exponential). If the outcomes are symmetrical around the mean, then +ve and –ve deviations are balanced, and skewness is 0. Graphical display of skewness Positive skewness: tail going right Negative skewness: tail going left Skewness and language By Zipf’s law (next week), word frequencies do not cluster around the mean. There are a few highly frequent words (making up a large proportion of overall word frequency) There are many highly infrequent words (f = 1 or f = 2) So the Zipfian distribution is highly skewed. We will hear more on the Zipfian distribution in the next lecture. The concept of information What is information? Main ingredient: an information source, which “transmits” symbols from a finite alphabet S every symbol is denoted si we call a sequence of such symbols a text assume a probability distribution s.t. every si has probability p(si) Example: a dice is an information source; every throw yields a symbol from the alphabet {1,2,3,4,5,6} 6 successive throws yield a text of 6 symbols Quantifying information Intuition: the more probable a symbol is, the less information it yields “something seen very often is not very surprising” So information is the inverse probability of the symbol I ( si ) log b 1 log p( si ) p ( si ) for some b > 1. Usually we use base 2 Another term for I(s) is surprisal Properties of I 1. Non-negative 2. If p(s) = 1, I(s) = 0 3. If 2 events s1, s2 are independent, then: I (si , s2 ) I ( p(s1 ) p( s2 )) I (s1 ) I ( s2 ) 4. Monotonic: slight changes in probability result in slight changes in I Aggregate measure of information What is the information content of a text (sequence of symbols)? 1. this is the same as finding the average information of a random variable the measure is called Entropy, denoted H Define X as a random variable over the symbols in our alphabet P(s) = P(X=s) for all s in our alphabet 2. Estimate H(P) Entropy The entropy (or information) of a probability distribution is 1 H b ( P) P( s) log b P( s) log b P( s) P( s ) sS sS entropy is the expected value (mean) of the surprisal the value is interpreted as the number of “bits” of information Entropy example Source = an 8-sided die Alphabet S = {1,2,3,4,5,6,7,8} every si has p = 1/8 8 8 1 1 1 H ( P) p( si ) log p ( si ) log log log 8 3 8 8 i 1 i 1 8 Interpretation of entropy The information contained in the distribution P (the more unpredictable the outcomes, the higher the entropy) The message length if the message was generated according to P and coded optimally Interpretation cont/d For the 8-sided die example, the result H(P)=3 tells us we need 3 bits on average to “transmit” the result of rolling an 8-sided die: 1 2 3 4 5 6 7 8 001 010 011 100 101 110 111 000 We can’t do it in less than 3 bits Entropy for multiple variables So far we have dealt with a single random variable The joint entropy of a pair of RVs: 1 H ( X , Y ) P ( x, y ) log b P ( x, y ) x X yY P( x, y ) log b P ( x, y ) x X yY Conditional Entropy Given X and Y, how much information about Y do we gain if we know X? a version of entropy using conditional probability: H(Y|X) H (Y | X ) P( x) H (Y | X x X x) P ( x) P ( y | x) log P ( y | x) x X yY P ( x) P ( y | x) log P ( y | x) x X yY Mutual information Mutual information Just as probability can change based on posterior knowledge, so can information. Suppose our distribution gives us the probability P(a) of observing the symbol a. Suppose we first observe the symbol b. If a and b are not independent, this should alter our information state with respect to the probability of observing a. i.e. we can compute p(a|b) Mutual info between two symbols The change in our information about a on observing b is: 1 1 log I (a; b) log P(a) P ( a | b) P ( a | b) log P(a) If a and b are completely independent, I(a;b)=0. Averaging mutual information We want to average mutual information between all values of a random variable A and those of a random variable B. P ( a | b) I ( a ; b) I ( A; b) i i i i P ( ai | b ) P(ai | b) log P ( ai ) And similarly: I (a; B) j P(bi | a) P(a | b j ) log P(b j ) Combining the two… I ( A; B) P(a ) I (a ; B) i i i P ( ai , b j ) P(ai , b j ) log P(ai ) P(b j ) i j I ( B; A) Thus, mutual info involves taking the joint probability and dividing by the individual probabilities. I.e. a comparison of the likelihood of observing a, b together vs. separately. Mutual Information: summary Gives a measure of reduction in uncertainty about a random variable X, given knowledge of Y quantifies how much information about X is contained in Y Some more on I(X;Y) In statistical NLP, we often calculate pointwise mutual information this is the mutual information between two points on a distribution I(x;y) rather than I(X;Y) used for some applications in lexical acquisition Mutual Information -- example Suppose we’re interested in the collocational strength of two words x and y e.g. bread and butter mutual information quantifies the likelihood of observing x and y together (in some window) If there is no interesting relationship, knowing about bread tells us nothing about the likelihood of encountering butter Here, P(x,y) = P(x)P(y) and I(x;y) = 0 This is the Church and Hanks (1991) approach. NB. The approach uses pointwise MI