NLP Project Full Cycle
5 likes12,152 views
The document outlines a comprehensive overview of Natural Language Processing (NLP) from a practitioner's perspective, detailing various aspects such as data types, common problems, and methodologies, including both rule-based and machine learning approaches. It discusses specific tasks like text language identification, tokenization, and parsing, alongside insights into data sourcing and model training. Additionally, it emphasizes the importance of experimental rigor in developing effective NLP systems while highlighting ongoing challenges and considerations in the field.
![linguist [noun]
1. A specialist in linguistics](https://image.slidesharecdn.com/lang-detect-161011092815/85/NLP-Project-Full-Cycle-8-320.jpg)
![linguist [noun]
1. A specialist in linguistics
linguistics [noun]
1. The scientific study of
language.](https://image.slidesharecdn.com/lang-detect-161011092815/85/NLP-Project-Full-Cycle-9-320.jpg)
![engineer [noun]
5. A person skilled in the
design and programming of
computer systems](https://image.slidesharecdn.com/lang-detect-161011092815/85/NLP-Project-Full-Cycle-17-320.jpg)
![Regular Expressions
Simplest regex: [^s]+
More advanced regex:
w+|[!"#$%&'*+,./:;<=>?@^`~…() {}[|]⟨⟩ ‒–—
«»“”‘’-]―
Even more advanced regex:
[+-]?[0-9](?:[0-9,.]*[0-9])?
|[w@](?:[w'’`@-][w']|[w'][w@'’`-])*[w']?
|["#$%&*+,/:;<=>@^`~…() {}[|] «»“”‘’']⟨⟩ ‒–—―
|[.!?]+
|-+
In fact, it works:
https://github.com/lang-uk/ner-uk/blob/master/doc
/tokenization.md](https://image.slidesharecdn.com/lang-detect-161011092815/85/NLP-Project-Full-Cycle-19-320.jpg)
![researcher [noun]
1. One who researches](https://image.slidesharecdn.com/lang-detect-161011092815/85/NLP-Project-Full-Cycle-23-320.jpg)
![researcher [noun]
1. One who researches
research [noun]
1. Diligent inquiry or
examination to seek or revise
facts, principles, theories,
applications, etc.; laborious
or continued search after
truth](https://image.slidesharecdn.com/lang-detect-161011092815/85/NLP-Project-Full-Cycle-24-320.jpg)
![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)](https://image.slidesharecdn.com/lang-detect-161011092815/85/NLP-Project-Full-Cycle-38-320.jpg)
![ML-based Parsing
The 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](https://image.slidesharecdn.com/lang-detect-161011092815/85/NLP-Project-Full-Cycle-39-320.jpg)
NLP Project Full Cycle
- 1. NLP Project Full Cycle Vsevolod Dyomkin 10/2016
- 2. A Bit about Me * Lisp programmer * 5+ years of NLP work at Grammarly * Occasional lecturer https://vseloved.github.io
- 3. Plan * Overview of NLP * NLP Data * Common NLP problems and approaches * Example NLP application: text language identification
- 4. What Is NLP? Transforming free-form text into structured data and back
- 5. What Is NLP? Transforming free-form text into structured data and back Intersection of: * Computational Linguistics * CompSci & AI * ML, Stats, Information Theory
- 6. Natural Language * ambiguous * noisy * evolving
- 7. Roles
- 8. linguist [noun] 1. A specialist in linguistics
- 9. linguist [noun] 1. A specialist in linguistics linguistics [noun] 1. The scientific study of language.
- 12. NLP Data Types of text data: * structured * semi-structured * unstructured “Data is ten times more powerful than algorithms.” -- Peter Norvig The Unreasonable Effectiveness of Data. http://youtu.be/yvDCzhbjYWs
- 13. Kinds of Data * Dictionaries * Databases/Ontologies * Corpora * Internet/user Data
- 14. 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, ...
- 15. Create Your Own! * Linguists * Crowdsourcing * By-product -- Johnatahn Zittrain http://goo.gl/hs4qB
- 16. Classic NLP Problems * Linguistically-motivated: segmentation, tagging, parsing * Analytical: classification, sentiment analysis * Transformation: translation, correction, generation * Conversation: question answering, dialog
- 17. engineer [noun] 5. A person skilled in the design and programming of computer systems
- 18. Tokenization Example: 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.
- 19. Regular Expressions Simplest regex: [^s]+ More advanced regex: w+|[!"#$%&'*+,./:;<=>?@^`~…() {}[|]⟨⟩ ‒–— «»“”‘’-]― Even more advanced regex: [+-]?[0-9](?:[0-9,.]*[0-9])? |[w@](?:[w'’`@-][w']|[w'][w@'’`-])*[w']? |["#$%&*+,/:;<=>@^`~…() {}[|] «»“”‘’']⟨⟩ ‒–—― |[.!?]+ |-+ In fact, it works: https://github.com/lang-uk/ner-uk/blob/master/doc /tokenization.md
- 20. 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
- 21. Rule-based NLP tools * SpamAssasin * LanguageTool * ELIZA * GATE
- 23. researcher [noun] 1. One who researches
- 24. researcher [noun] 1. One who researches research [noun] 1. Diligent inquiry or examination to seek or revise facts, principles, theories, applications, etc.; laborious or continued search after truth
- 25. Models
- 26. Statistical Approach “Probability theory is nothing but common sense reduced to calculation.” -- Pierre-Simon Laplace
- 27. Language Models Question: what is the probability of a sequence of words/sentence?
- 28. 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 …
- 29. Ngrams Apply 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)
- 30. Spam Filtering A 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?
- 31. Bag-of-words Model * 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
- 32. Bag-of-words Model * 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 http://www.paulgraham.com/spam.html - A Plan for Spam Initial results: recall: 92%, precision: 98.84% Improved results: recall: 99.5%, precision: 99.97%
- 33. 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)
- 35. 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-fas t-algorithm-for-natural-language-dependency-parsing/
- 38. 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)
- 39. ML-based Parsing The 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
- 40. The Hierarchy of ML Models Linear: * (Averaged) Perceptron * Maximum Entropy / LogLinear / Logistic Regression; Conditional Random Field * SVM Non-linear: * Decision Trees, Random Forests, Boosted Trees * Artificial Neural networks
- 41. Semantics Question: how to model relationships between words?
- 42. Semantics Question: how to model relationships between words? Answer: build a graph Wordnet Freebase DBPedia
- 43. Word Similarity Next question: now, how do we measure those relations?
- 44. Word Similarity Next question: now, how do we measure those relations? * different Wordnet similarity measures
- 45. 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))
- 46. Distributional Semantics Distributional hypothesis: "You shall know a word by the 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 repr (Brown clusters)
- 47. Steps to Develop an 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 * Setup and perform a proper experiment (series of experiments) * Optimize the system for production
- 48. Going into Prod * NLP tasks are usually CPU-intensive but stateless * General-purpose NLP frameworks are (mostly) not production-ready * Don't trust research results * Value pre- and post- processing * Gather user feedback
- 49. Text Language Identification Not an unsolved problem: * https://github.com/CLD2Owners/cld2 - C++ * https://github.com/saffsd/langid.py - Python * https://github.com/shuyo/language-detection/ - Java To read: https://blog.twitter.com/2015/evaluating-language-identifi cation-performance http://blog.mikemccandless.com/2011/10/accuracy-and-perfor mance-of-googles.html http://lab.hypotheses.org/1083 http://labs.translated.net/language-identifier/
- 50. WILD Challenges
- 52. YALI WILD * All of them use weak models * Wanted to use Wiktionary — 150+ languages, always evolving * Wanted to do in Lisp
- 53. WILD Linguistics * Scripts vs languages http://www.omniglot.com/writing/langalph.htm * Languages distribution https://en.wikipedia.org/wiki/Languages_used_o n_the_Internet#Content_languages_for_websites * Frequency word lists https://invokeit.wordpress.com/frequency-word- lists/ * Word segmentation?
- 54. WILD Data Wiktionary Wikipedia data: used abstracts, ~175 languages - download & store - process (SAX parsing) - setup learning & test data sets 10,778,404 unique words 481,581 unique character trigrams
- 55. WILD Engineering * Initial model size ~1G - script hacks & Huffman coding to the rescue * Model pruning * Proper probability calculations * Efficient testing * Properly saving the model * Library & public API