crash course in natural language processing (2016)

Crash Course inNatural Language Processing

Vsevolod Dyomkin04/2016

A Bit about Me

* Lisp programmer* 5+ years of NLP work at Grammarly * Occasional lecturer

https://vseloved.github.io

A Bit about GrammarlyThe best English language writing app

Spellcheck - Grammar check - Style improvement - Synonyms and word choice Plagiarism check

Plan* Overview of NLP* Where to get Data* Common NLP problems and approaches* How to develop an NLP system

What Is NLP?Transforming free-form text into structured data and back

What Is NLP?Transforming free-form text into structured data and back

Intersection of:* Computational Linguistics* CompSci & AI* Stats & Information Theory

Linguistic Basis

* Syntax (form)* Semantics (meaning)* Pragmatics (intent/logic)

Natural Language

* ambiguous* noisy* evolving

Time flies like an arrow.Fruit flies like a banana.

I read a story about evolution in ten minutes.I read a story about evolution in the last million years.

NLP & DataTypes of text data:* structured* semi-structured* unstructured

“Data is ten times more

powerful than algorithms.”-- Peter NorvigThe Unreasonable Effectiveness of Data.http://youtu.be/yvDCzhbjYWs

Kinds of Data* Dictionaries* Databases/Ontologies* Corpora* User Data

Where to Get Data?* Linguistic Data Consortium http://www.ldc.upenn.edu/ * Common Crawl* Wikimedia* Wordnet* APIs: Twitter, Wordnik, ...* University sites & the academic community: Stanford, Oxford, CMU, ...

Create Your Own!* Linguists* Crowdsourcing* By-product

-- Johnatahn Zittrain http://goo.gl/hs4qB

Classic NLP Problems* Linguistically-motivated: segmentation, tagging, parsing

* Analytical: classification, sentiment analysis

* Transformation: translation, correction, generation

* Conversation:question answering, dialog

TokenizationExample:This is a test that isn't so simple: 1.23."This" "is" "a" "test" "that" "is" "n't" "so" "simple" ":" "1.23" "."

Issues:* Finland’s capital - Finland Finlands Finland’s* what’re, I’m, isn’t - what ’re, I ’m, is n’t* Hewlett-Packard or Hewlett Packard * San Francisco - one token or two?* m.p.h., PhD.

Regular ExpressionsSimplest regex: [^\s]+

More advanced regex:

\w+|[!"#$%&'*+,\./:;<=>?@^`~…\(\) {}\[\|\]⟨⟩ ‒–—«»“”‘’-]―

Even more advanced regex:

[+-]?[0-9](?:[0-9,\.]*[0-9])?|[\w@](?:[\w'’`@-][\w']|[\w'][\w@'’`-])*[\w']?|["#$%&*+,/:;<=>@^`~…\(\) {}\[\|\] «»“”‘’']⟨⟩ ‒–—―|[\.!?]+|-+

Post-processing

* concatenate abbreviations and decimals* split contractions with regexes 2-character: i['‘’`]m|(?:s?he|it)['‘’`]s|(?:i|you|s?he|we|they) ['‘’`]d$

3-character: (?:i|you|s?he|we|they)['‘’`](?:ll|[vr]e)|n['‘’`]t$

Rule-based Approach* easy to understand and reason about* can be arbitrarily precise* iterative, can be used to gather more data

Limitations:* recall problems* poor adaptability

Rule-based NLP tools

* SpamAssasin* LanguageTool* ELIZA* GATE

Statistical Approach

“Probability theoryis nothing butcommon sensereduced to calculation.”-- Pierre-Simon Laplace

Language Models

Question: what is the probability of a sequence of words/sentence?

Language Models

Question: what is the probability of a sequence of words/sentence?

Answer: Apply the chain rule

P(S) = P(w0) * P(w1|w0) * P(w2|w0 w1) * P(w3|w0 w1 w2) * …

where S = w0 w1 w2 …

NgramsApply Markov assumption: each word depends only on N previous words (in practice N=1..4 which results in bigrams-fivegrams, because we include the current word also).

If n=2: P(S) = P(w0) * P(w1|w0) * P(w2|w0 w1) * P(w3|w1 w2) * …

According to the chain rule:

P(w2|w0 w1) = P(w0 w1 w2) / P(w0 w1)

Spelling Correction

Problem: given an out-of-dictionary word return a list of most probable in-dictionary corrections.

http://norvig.com/spell-correct.html

Edit DistanceMinimum-edit (Levenstein) distance the –minimum number of insertions/deletions/substitutions needed to transform string A into B.

Other distance metrics:* the Damerau-Levenstein distance adds another operation: transposition* the longest common subsequence (LCS) metric allows only insertion and deletion, not substitution* the Hamming distance allows only substitution, hence, it only applies to strings of the same length

Dynamic ProgrammingInitialization: D(i,0) = i D(0,j) = j

Recurrence relation: For each i = 1..M For each j = 1..N D(i,j) = D(i-1,j-1), if X(i) = Y(j)

otherwise: min D(i-1,j) + w_del(Y(j)) D(i,j-1) + w_ins(X(i)) D(i-1,j-1) + w_subst(X(i),Y(j))

Noisy Channel ModelGiven an alphabet A, let A* be the set of all finite strings over A. Let the dictionary D of valid words be some subset of A*.

The noisy channel is the matrix G = P(s|w) where w in D is the intended word and s in A* is the scrambled word that was actually received.

P(s|w) = sum(P(x(i)|y(i))) for x(i) in s* (s aligned with w) for y(i) in w* (w aligned with s)

Machine Learning Approach

Spam FilteringA 2-class classification problem with a bias towards minimizing FPs.

Default approach: rule-based (SpamAssassin)

Problems:* scales poorly* hard to reach arbitrary precision* hard to rank the importance of complex features?

Bag-of-words Models* each word is a feature* each word is independent of others* position of the word in a sentence is irrelevant

Pros:* simple* fast* scalable

Limitations:* independence assumption doesn't hold

Initial results: recall: 92%, precision: 98.84% Improved results: recall: 99.5%, precision: 99.97%

http://www.paulgraham.com/spam.html

Naive Bayes Classifier

P(Y|X) = P(Y) * P(X|Y) / P(X)select Y = argmax P(Y|x) Naive step:

P(Y|x) = P(Y) * prod(P(x|Y)) for all x in X

(P(x) is marginalized out because it's the same for all Y)

Dependency Parsing

nsubj(ate-2, They-1)root(ROOT-0, ate-2)det(pizza-4, the-3)dobj(ate-2, pizza-4)prep(ate-2, with-5)pobj(with-5, anchovies-6)

https://honnibal.wordpress.com/2013/12/18/a-simple-fast-algorithm-for-natural-language-dependency-parsing/

Shift-reduce Parsing

Shift-reduce Parsing

ML-based ParsingThe parser starts with an empty stack, and a buffer index at 0, with no dependencies recorded. It chooses one of the valid actions, and applies it to the state. It continues choosing actions and applying them until the stack is empty and the buffer index is at the end of the input.

SHIFT = 0; RIGHT = 1; LEFT = 2 MOVES = [SHIFT, RIGHT, LEFT]

def parse(words, tags): n = len(words) deps = init_deps(n) idx = 1 stack = [0] while stack or idx < n: features = extract_features(words, tags, idx, n, stack, deps) scores = score(features) valid_moves = get_valid_moves(i, n, len(stack)) next_move = max(valid_moves, key=lambda move: scores[move]) idx = transition(next_move, idx, stack, parse) return tags, parse

Averaged Perceptron

def train(model, number_iter, examples): for i in range(number_iter): for features, true_tag in examples: guess = model.predict(features) if guess != true_tag: for f in features: model.weights[f][true_tag] += 1 model.weights[f][guess] -= 1 random.shuffle(examples)

Features* Word and tag unigrams, bigrams, trigrams* The first three words of the buffer* The top three words of the stack* The two leftmost children of the top of the stack* The two rightmost children of the top of the stack* The two leftmost children of the first word in the buffer* Distance between top of buffer and stack

Discriminative ML Models

Linear:* (Averaged) Perceptron* Maximum Entropy / LogLinear / Logistic Regression; Conditional Random Field* SVM

Non-linear:* Decision Trees, Random Forests* Other ensemble classifiers* Neural networks

SemanticsQuestion: how to model relationships between words?

SemanticsQuestion: how to model relationships between words?Answer: build a graph

WordnetFreebaseDBPedia

Word Similarity

Next question: now, how do we measure those relations?

Word Similarity

Next question: now, how do we measure those relations?

* different Wordnet similarity measures

Word Similarity

Next question: now, how do we measure those relations?

* different Wordnet similarity measures

* PMI(x,y) = log(p(x,y) / p(x) * p(y))

Distributional Semantics

Distributional hypothesis:"You shall know a word bythe company it keeps"--John Rupert Firth

Word representations:* Explicit representation Number of nonzero dimensions: max:474234, min:3, mean:1595, median:415* Dense representation (word2vec, GloVe)* Hierarchical representation (Brown clustering)

Steps to Developan NLP System

* Translate real-world requirements into a measurable goal* Find a suitable level and representation* Find initial data for experiments* Find and utilize existing tools and Frameworks where possible* Don't trust research results* Setup and perform a proper experiment (series of experiments)

Going into Prod

* NLP tasks are usually CPU-intensive but stateless * General-purpose NLP frameworks are (mostly) not production-ready* Value pre- and post- processing* Gather user feedback

Final WordsWe have discussed:* linguistic basis of NLP

- although some people manage to do NLP without it:http://arxiv.org/pdf/1103.0398.pdf

* rule-based & statistical/ML approaches* different concrete tasks

We haven't covered:* all the different tasks, such as MT,

question answering, etc. (but they use the same technics)* deep learning for NLP* natural language understanding (which remains an unsolved problem)