/
CSC 594 Topics in AI – Natural Language Processing Spring CSC 594 Topics in AI – Natural Language Processing Spring

CSC 594 Topics in AI – Natural Language Processing Spring - PowerPoint Presentation

jane-oiler
jane-oiler . @jane-oiler
Follow
347 views
Uploaded On 2019-11-01

CSC 594 Topics in AI – Natural Language Processing Spring - PPT Presentation

CSC 594 Topics in AI Natural Language Processing Spring 2018 6 Language Models Some slides adapted from Jurafsky amp Martin Word Prediction Guess the next word So I notice three guys standing on the ID: 761680

jurafsky language martin processing language jurafsky processing martin speech bigram probability counts model

Share:

Link:

Embed:

Download Presentation from below link

Download Presentation The PPT/PDF document "CSC 594 Topics in AI – Natural Languag..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.


Presentation Transcript

CSC 594 Topics in AI –Natural Language Processing Spring 20186. Language Models(Some slides adapted from Jurafsky & Martin)

Word Prediction Guess the next word...So I notice three guys standing on the ???What are some of the knowledge sources you used to come up with those predictions?Speech and Language Processing - Jurafsky and Martin 2

Word Prediction We can formalize this task using what are called N-gram modelsN-grams are token sequences of length N- Our earlier example contains the following 2-grams (aka bigrams) (So I), (I notice), (notice three), (three guys), (guys standing), (standing on), (on the) Given knowledge of counts of N-grams such as these, we can guess likely next words in a sequence. Speech and Language Processing - Jurafsky and Martin 3

N-Gram Models More formally, we can use knowledge of the counts of N-grams to assess the conditional probability of candidate words as the next word in a sequence.Or, we can use them to assess the probability of an entire sequence of words.Pretty much the same thing as we’ll see... Speech and Language Processing - Jurafsky and Martin 4

Applications It turns out that being able to assess the probability of a sequence is an extremely useful thing to be able to do.As we’ll see, it lies at the core of many applications Automatic speech recognition Handwriting and character recognition Spam detection Sentiment analysis Spelling correction Machine translation … Speech and Language Processing - Jurafsky and Martin 5

Counting Simple counting lies at the core of any probabilistic approach. So let’s first take a look at what we’re counting.He stepped out into the hall, was delighted to encounter a water brother.13 tokens, 15 if we include “,” and “ . ” as separate tokens. Assuming we include the comma and period as tokens, how many bigrams are there? Speech and Language Processing - Jurafsky and Martin 6

Counting: Types and Tokens How aboutThey picnicked by the pool, then lay back on the grass and looked at the stars.18 tokens (again counting punctuation)But we might also note that “the ” is used 3 times, so there are only 16 unique types (as opposed to tokens). In going forward, we’ll have occasion to focus on counting both types and tokens of both words and N -grams . When we’re looking at isolated words we’ll refer to them as unigrams Speech and Language Processing - Jurafsky and Martin 7

Language Modeling Now that we know how to count, back to word predictionWe can model the word prediction task as the ability to assess the conditional probability of a word given the previous words in the sequence P(wn|w1,w2…wn-1) We’ll call a statistical model that can assess this a Language Model Speech and Language Processing - Jurafsky and Martin 8

Language Modeling How might we go about calculating such a conditional probability? One way is to use the definition of conditional probabilities and look for counts. So to getP(the | its water is so transparent that)By definition that’sP(its water is so transparent that the) P(its water is so transparent that) We can get each of those from counts in a large corpus. Speech and Language Processing - Jurafsky and Martin 9

Very Easy Estimate How to estimate?P(the | its water is so transparent that)P(the | its water is so transparent that) =Count(its water is so transparent that the) Count(its water is so transparent that) Speech and Language Processing - Jurafsky and Martin 10

Very Easy Estimate According to Google those counts are 12000 and 19000 so the conditional probability of interest is... P(the | its water is so transparent that) = 0.63 Speech and Language Processing - Jurafsky and Martin 11

Language Modeling Unfortunately, for most sequences and for most text collections we won’t get good estimates from this method.What we’re likely to get is 0. Or worse 0/0.Clearly, we’ll have to be a little more clever. Let’s first use the chain rule of probability And then apply a particularly useful independence assumption Speech and Language Processing - Jurafsky and Martin 12

The Chain Rule Recall the definition of conditional probabilitiesRewriting:For sequences... P(A,B,C,D) = P(A)P(B|A)P(C|A,B)P(D|A,B,C) In general P(x 1 ,x 2 ,x 3 ,…xn ) = P(x 1 )P(x 2 |x 1 )P(x 3|x 1 ,x 2 )…P(x n |x 1 …x n-1 ) Speech and Language Processing - Jurafsky and Martin 13

The Chain Rule P(its water was so transparent)=P(its)* P(water|its)* P(was|its water)* P(so|its water was)* P(transparent|its water was so) Speech and Language Processing - Jurafsky and Martin 14

Unfortunately There are still a lot of possible sequences in thereIn general, we’ll never be able to get enough data to compute the statistics for those longer prefixesSame problem we had for the strings themselves Speech and Language Processing - Jurafsky and Martin 15

Independence Assumption Make the simplifying assumptionP(lizard|the,other,day,I,was,walking,along,and,saw,a) = P(lizard|a)Or maybeP(lizard|the,other,day,I,was,walking,along,and,saw,a ) = P( lizard|saw,a ) That is, the probability in question is to some degree independent of its earlier history. Speech and Language Processing - Jurafsky and Martin 16

Independence Assumption This particular kind of independence assumption is called a Markov assumption after the Russian mathematician Andrei Markov. Speech and Language Processing - Jurafsky and Martin 17

So for each component in the product replace with the approximation (assuming a prefix of N - 1) Bigram version Markov Assumption Speech and Language Processing - Jurafsky and Martin 18

Estimating Bigram Probabilities The Maximum Likelihood Estimate (MLE) Speech and Language Processing - Jurafsky and Martin 19

An Example <s> I am Sam </s><s> Sam I am </s><s> I do not like green eggs and ham </s> Speech and Language Processing - Jurafsky and Martin 20

Berkeley Restaurant Project Sentences can you tell me about any good cantonese restaurants close bymid priced thai food is what i’m looking fortell me about chez panisse can you give me a listing of the kinds of food that are available i ’ m looking for a good place to eat breakfast when is caffe venezia open during the day Speech and Language Processing - Jurafsky and Martin 21

Bigram Counts Vocabulary size is 1446 |V|Out of 9222 sentencesEg. “I want” occurred 827 times Speech and Language Processing - Jurafsky and Martin 22

Bigram Probabilities Divide bigram counts by prefix unigram counts to get bigram probabilities. Speech and Language Processing - Jurafsky and Martin 23

Bigram Estimates of Sentence Probabilities P(<s> I want english food </s>) = P(i|<s>)* P( want|I )* P( english|want )* P( food|english )* P(</s>|food)* =.000031 Speech and Language Processing - Jurafsky and Martin 24

Kinds of Knowledge P(english|want) = .0011P(chinese|want) = .0065 P( to|want ) = .66 P(eat | to) = .28 P(food | to) = 0 P(want | spend) = 0 P ( i | <s>) = .25 As crude as they are, N -gram probabilities capture a range of interesting facts about language. World knowledge Syntax Discourse Speech and Language Processing - Jurafsky and Martin 25

Shannon’ s MethodAssigning probabilities to sentences is all well and good, but it’s not terribly illuminating. A more entertaining task is to turn the model around and use it to generate random sentences that are like the sentences from which the model was derived. Generally attributed to Claude Shannon. Speech and Language Processing - Jurafsky and Martin 26

Shannon’ s MethodSample a random bigram (<s>, w) according to the probability distribution over bigramsNow sample a new random bigram (w, x) according to its probabilityWhere the prefix w matches the suffix of the first bigram chosen. And so on until we randomly choose a (y, </s>) Then string the words together <s> I I want want to to eat eat Chinese Chinese food food </s> Speech and Language Processing - Jurafsky and Martin 27

Shakespeare Speech and Language Processing - Jurafsky and Martin 28

Shakespeare as a Corpus N=884,647 tokens, V=29,066Shakespeare produced 300,000 bigram types out of V2= 844 million possible bigrams... So, 99.96% of the possible bigrams were never seen (have zero entries in the table)This is the biggest problem in language modeling; we’ll come back to it. Quadrigrams are worse: What's coming out looks like Shakespeare because it is Shakespeare Speech and Language Processing - Jurafsky and Martin 29

The Wall Street Journal is Not Shakespeare Speech and Language Processing - Jurafsky and Martin 30

Model Evaluation How do we know if our models are any good?And in particular, how do we know if one model is better than another.Well Shannon’s game gives us an intuition.The generated texts from the higher order models sure sounds better . That is, they sound more like the text the model was obtained from. The generated texts from the WSJ and Shakespeare models look different That is, they look like they ’ re based on different underlying models. But what does that mean? Can we make that notion operational? Speech and Language Processing - Jurafsky and Martin 31

Evaluating N -Gram ModelsBest evaluation for a language modelPut model A into an applicationFor example, a machine translation systemEvaluate the performance of the application with model A Put model B into the application and evaluate Compare performance of the application with the two models Extrinsic evaluation Speech and Language Processing - Jurafsky and Martin 32

Evaluation Extrinsic evaluationThis is really time-consuming and hardNot something you want to do unless you’re pretty sure you’ve got a good solutionSo As a temporary solution, in order to run rapid experiments we evaluate N-grams with an intrinsic evaluation An evaluation that tries to capture how good the model is intrinsically, not how much it improves performance in some larger system Speech and Language Processing - Jurafsky and Martin 33

Evaluation Standard methodTrain parameters of our model on a training set.Evaluate the model on some new data: a test set. A dataset which is different than our training set, but is drawn from the same source Speech and Language Processing - Jurafsky and Martin 34

Perplexity The intuition behind perplexity as a measure is the notion of surprise.How surprised is the language model when it sees the test set?Where surprise is a measure of...Gee, I didn’t see that coming...The more surprised the model is, the lower the probability it assigned to the test set The higher the probability, the less surprised it was Speech and Language Processing - Jurafsky and Martin 35

Perplexity Perplexity is the probability of a test set (assigned by the language model), as normalized by the number of words:Chain rule: For bigrams: Minimizing perplexity is the same as maximizing probability The best language model is one that best predicts an unseen test set Speech and Language Processing - Jurafsky and Martin 36

Lower perplexity means a better model Training 38 million words, test 1.5 million words, WSJ Speech and Language Processing - Jurafsky and Martin 37

Practical Issues Once we start looking at test data, we’ll run into words that we haven’t seen before. So our models won’t wor Standard solution Given a corpus Create an unknown word token <UNK > Create a fixed lexicon L, of size V From a dictionary or A subset of terms from the training set At text normalization phase, any training word not in L is changed to <UNK> Collect counts for that as for any normal word At test time Use UNK counts for any word not seen in training Speech and Language Processing - Jurafsky and Martin 38

Practical Issues Multiplying a bunch of really small numbers < 0 is a really bad idea.Multiplication is slowAnd underflow is likelySo do everything in log spaceAvoid underflow(also adding is faster than multiplying) Speech and Language Processing - Jurafsky and Martin 39

Smoothing Back to ShakespeareRecall that Shakespeare produced 300,000 bigram types out of V2= 844 million possible bigrams... So, 99.96% of the possible bigrams were never seen (have zero entries in the table)Does that mean that any sentence that contains one of those bigrams should have a probability of 0?For generation (shannon game) it means we’ll never emit those bigrams But for analysis it’ s problematic because if we run across a new bigram in the future then we have no choice but to assign it a probability of zero. Speech and Language Processing - Jurafsky and Martin 40

Zero Counts Some of those zeros are really zeros... Things that really aren’t ever going to happenOn the other hand, some of them are just rare events. If the training corpus had been a little bigger they would have had a countWhat would that count be in all likelihood?Zipf ’ s Law (long tail phenomenon): A small number of events occur with high frequency A large number of events occur with low frequency You can quickly collect statistics on the high frequency events You might have to wait an arbitrarily long time to get valid statistics on low frequency events Result: Our estimates are sparse! We have no counts at all for the vast number of things we want to estimate! Answer: Estimate the likelihood of unseen (zero count) N-grams! Speech and Language Processing - Jurafsky and Martin 41

Laplace Smoothing Also called Add-One smoothingJust add one to all the counts!Very simpleMLE estimate:Laplace estimate: Speech and Language Processing - Jurafsky and Martin 42

Bigram Counts Speech and Language Processing - Jurafsky and Martin 43

Laplace-Smoothed Bigram Counts Speech and Language Processing - Jurafsky and Martin 44

Laplace-Smoothed Bigram Probabilities Speech and Language Processing - Jurafsky and Martin 45

Reconstituted Counts Speech and Language Processing - Jurafsky and Martin 46

Reconstituted Counts (2) Speech and Language Processing - Jurafsky and Martin 47

Big Change to the Counts! C(want to) went from 608 to 238!P(to|want) from .66 to .26!Discount d= c*/cd for “chinese food ” =.10!!! A 10x reduction So in general, Laplace is a blunt instrument Could use more fine-grained method (add-k) Because of this Laplace smoothing not often used for language models, as we have much better methods Despite its flaws Laplace (add -1) is still used to smooth other probabilistic models in NLP and IR, especially For pilot studies In document classification In domains where the number of zeros isn’t so huge. Speech and Language Processing - Jurafsky and Martin 48

Better Smoothing An intuition used by many smoothing algorithms is to use the count of things we’ve seen once to help estimate the count of things we’ve never seenSpeech and Language Processing - Jurafsky and Martin 49

Backoff and InterpolationSometimes it helps to use less contextCondition on less context for contexts you haven’t learned much about Backoff use trigram if you have good evidence, otherwise bigram, otherwise unigram Interpolation mix unigram, bigram, trigramInterpolation works better Speech and Language Processing - Jurafsky and Martin 50

Linear Interpolation Simple interpolationLambdas conditional on context Speech and Language Processing - Jurafsky and Martin 51

How to set the lambdas? Use a held-out corpusChoose λs to maximize the probability of held-out data:Fix the N-gram probabilities (using the training data) Then search for λs that give largest probability to held-out set. Training Data Held-Out Data Test Data Speech and Language Processing - Jurafsky and Martin 52

Types, Tokens and Fish Much of what’s coming up was first studied by biologists who are often faced with 2 related problems Determining how many species occupy a particular area (types) And determining how many individuals of a given species are living in a given area (tokens) Speech and Language Processing - Jurafsky and Martin 53

One Fish Two Fish Imagine you are fishingThere are 8 species: carp, perch, whitefish, trout, salmon, eel, catfish, bassNot sure where this fishing hole is...You have caught up to now10 carp, 3 perch, 2 whitefish, 1 trout, 1 salmon, 1 eel = 18 fish How likely is it that the next fish to be caught is an eel? How likely is it that the next fish caught will be a member of newly seen species? Now how likely is it that the next fish caught will be an eel? Speech and Language Processing - Jurafsky and Martin 54

Fish LessonWe need to steal part of the observed probability mass to give it to the as yet unseen N-Grams. Questions are: How much to stealHow to redistribute itSpeech and Language Processing - Jurafsky and Martin55

Absolute DiscountingJust subtract a fixed amount from all the observed counts (call that d). Redistribute it proportionally based on observed dataSpeech and Language Processing - Jurafsky and Martin56

Absolute Discounting w/ Interpolation d iscounted bigram unigram Interpolation weight Speech and Language Processing - Jurafsky and Martin 57

Better estimate for probabilities of lower-order unigrams! Shannon game: I can’t see without my reading___________?“Francisco” is more common than “glasses” … but “ Francisco” frequently follows “San” So P(w) isn’t what we want Francisco Kneser -Ney Smoothing g lasses Speech and Language Processing - Jurafsky and Martin 58

Kneser-Ney Smoothing Pcontinuation(w): “How likely is a word to appear as a novel continuation?For each word, count the number of bigram types it completes Speech and Language Processing - Jurafsky and Martin 59

Kneser -Ney SmoothingNormalize that by the total number of word bigram types to get a true probability Speech and Language Processing - Jurafsky and Martin 60

Kneser-Ney Smoothing λ is a normalizing constant; the probability mass we’ve discounted t he normalized discount The number of word types that can follow w i-1 Speech and Language Processing - Jurafsky and Martin 61

Bigram Counts Speech and Language Processing - Jurafsky and Martin 62

BERPLet’s look at “chinese food”. We’ll need:Count(“chinese food”)Count(“chinese”)P_continuation(“food”)Count of bigrams “food” completesCount of all bigram typesCount of bigrams that “chinese” starts Speech and Language Processing - Jurafsky and Martin 63

BERPLet’s look at “chinese food”. We’ll need:Count(“chinese food”) 82Count(“chinese”) 158P_continuation(“food”)Count of bigrams “food” completes 110Count of all bigram types 9421Count of bigrams that “chinese” starts 17 Speech and Language Processing - Jurafsky and Martin 64