Just Enough NLP with Python

Use NLTK to do a little bit of NLP with Python. Lots of code examples. Presented by Andrew Montalenti, CTO of Parse.ly. See http://parse.ly

Slides were created using reST and S5. You can read the slides in reST format (which is quite pleasant and lets you easily copy/paste code into IPython) here: https://raw.github.com/Parsely/python-nlp-slides/master/index.rst

You can also read the slides in your browser here:
http://bit.ly/nlp-slides

1. Just Enough NLP with Python Author:Andrew Montalenti Date: 2012-10-26 Just Enough NLP with Python
2. Meta Information Me: I've been using Python for 10 years. I use Python full-time, and have for the last 3 years. Startup: I'm co-founder/CTO of Parse.ly ❏, a tech startup in the digital media space. E-mail me: andrew@parsely.com ❏ Follow me on Twitter: @amontalenti ❏ Connect on LinkedIn: http://linkedin.com/in/andrewmontalenti ❏
3. Parse.ly What do we do? How do we do it?
4. Complex Media Male-focused media company started by Mark Ecko, American fashion designer and entrepreneur. Huge monthly traffic numbers, lots of readers across the world. Early adopters of Parse.ly.
5. Eye Disease Progressive outer retinal necrosis. Also known as Varicella zoster virus retinitis (VZVR), it is an aggressive, necrotizing inflammation of the eye's retina caused by herpes varicella zoster virus.
6. So, a question Why was I researching eye disease on a Saturday to debug a customer problem?
7. Answer Progressive Outer Retinal Necrosis
8. ANSWER P O R N
9. Lesson Learned Customer Metadata is Hard (See my talk on crawling/metadata later today) Topic Ontologies are Hard (See Didier's Wikipedia talk later) Thorny problems in NLP persist: disambiguation, overfitting NLP may not be the answer; consider IR Parse.ly is even undergoing an NLP -> IR shift in approach That said, NLTK keeps getting better
10. Meta Slide http://bit.ly/nlp-slides ❏ reST, S5, Python all the way down
11. NLTK Hello, World >>> import nltk >>> msg = "Hello, World!" >>> nltk.wordpunct_tokenize(msg) ['Hello', ',', 'World', '!']
12. Why is NLTK a Pythonic library? >>> len(dir(nltk)) 355 >>> fd = inspect_module(nltk) >>> fd.items() [('class', 172), ('function', 107), ('module', 48), ('other', 28)]
13. Here's inspect_module for reference import nltk import inspect def inspect_module(module): fd = nltk.FreqDist() inspections = ["function", "module", "class"] for item in vars(module).itervalues(): matched = False for inspection in inspections: if getattr(inspect, "is%s" % inspection)(item): matched = True fd.inc(inspection) if not matched: fd.inc("other") return fd
14. Batteries Included With more than 100 classes and 100 functions in the root nltk module, nltk certain adheres to "flat is better than nested" However, in the best Pythonic style, the flattened namespace is also neatly organized into nearly 48 submodules, of which many contain sub-sub-modules. Therefore, it also respects "namespaces are one honking great idea"
15. Practicality beats purity Although NLTK is, in name, a "natural language toolkit", it also includes some generally useful modules that are notably missing from Python Stdlib. I've already used one of these, nltk.FreqDist, which is a generic and Python "frequency distribution" class. It is actually dict-like, which means it supports the full dictionary protocol, but adds a few more functions. The keys are labels and values are integers representing number of occurrences of each supplied label. You increment labels with fd.inc().
16. Practicality: HTML cleaning nltk.clean_html is a nice HTML-stripping function. >>> nltk.clean_html(""" <p>This is some article text with <a href='http://google.com'> a link to Google</a></p>""") 'This is some article text with a link to Google'
17. Practicality: Tree data structure nltk.Tree is a powerful abstraction for grouping trees and subtrees. Each Tree contains leaves and subtrees. Leaves are simply literal values, but subtrees are treated specially. Further, trees can have arbitrary node properties. Finally, Tree instances can be constructed either using Python procedural code or a simple text DSL implemented by the Tree.parse class method.
18. Syntax parse tree example John hit the ball. JOHN hit the ball. john HIT the ball. john hit THE BALL.
19. Syntax parse tree example
21. Syntax parse tree example
23. Tree example: parsing >>> Tree.parse(""" (S (PERSON Herman Cain) runs for (POSITION president of the (COUNTRY United States) ) )""") Tree('S', [Tree('PERSON', ['Herman', 'Cain']), 'runs', ...
24. Tree example: printing >>> print tree.pprint(margin=40, nodesep=" ->", parens=["", ""]) S -> PERSON -> Herman Cain runs for POSITION -> president of the COUNTRY -> United States
25. Tree example: drawing! >>> tree.draw()
26. A taste of what's to come >>> tree = entities("""Hermain Cain runs for president of the United States""") >>> print tree.pprint(margin=40, nodesep=" ->", parens=["", ""]) S -> NE -> Herman/NNP Cain/NNP runs/VBZ president/NN of/IN the/DT NE -> United/NNP States/NNPS
27. Text and TextCollection One last set of data structures to be aware of are implemented in the nltk.Text and nltk.TextCollection classes. A Text is nothing more than an in-memory data structure of a variety of a collection of tokens, with the ability to do quick text analyses such as term frequency, collocation, similarity, and simple regex-based searching. A TextCollection is a grouping of Text instances that allows you to do corpus-wide calculations (such as term frequency, inverse document frequency, and yes, tf/idf!)
28. Text example >>> t1 = nltk.Text(nltk.word_tokenize(""" Barack Obama is president of the United States. Mr. Obama was elected in 2008.""")) >>> t1.count("Obama") 2
29. TextCollection example (1) >>> t2 = nltk.Text(nltk.word_tokenize(""" Barack Obama is giving a speech on Iraq tomorrow""")) >>> t3 = nltk.Text(nltk.word_tokenize(""" Barack Obama's speech illustrates the president's goal to leave Iraq""")) >>> col = nltk.TextCollection([t1, t2, t3])
30. TextCollection example (2) >>> col.vocab().items()[0:4] [('Obama', 4), ('Barack', 3), ("'s", 2), ('Iraq', 2)] >>> col.tf("Barack", t1) 0.066666666666666666 >>> col.tf("Obama", t1) 0.13333333333333333 >>> col.idf("Obama") 0.0 >>> col.idf("Iraq") 0.40546510810816438 >>> col.tf_idf("Obama", t1) 0.0 >>> col.tf_idf("Iraq", t2) 0.045051678678684932 >>> col.collocations() Barack Obama
31. Quick note on Text/TextCollection Though these classes are good for illustration purposes, I find this to be one of the less polished parts of NLTK. For more formal support for texts and text collections, one should use Solr in production. I've considered experimenting with Whoosh (basically, "a Solr in Python," but simpler/less scalable) but never found a good reason to avoid simply loading text documents into Solr.
32. Rule vs. Data-based Corpus Linguistics Part of the principle behind NLTK is that 100% rule-based language processing has failed to produce the results necessary for large-scale NLP needs. NLTK's approach is to take the best of the rule-based world (parse trees, syntactic decomposition, tagging) and combine it with the lessons learned by the information retrieval community. That is, often data can inform models better than cleverness.
33. Practicality wins again: nltk.data The nltk.data module offers access to a slew of off-the-shelf models that are widely used in academia, and is extensible so that you can add your own. The data tends to be stored in high-speed disk indexes (e.g. cPickle files) so that performance is acceptable as long as fast I/O is available. import nltk nltk.download()
34. On the NLTK menu (1) So, we have seen that NLTK provides some basic utilities that will likely make NLP easier, such as trees and statistical data structures. What else does NLTK offer? Much more than you might expect. Here are some highlights: • nltk.tokenize: a variety of tokenizers using fast, rule-based algorithms. These are familiar to users of Lucene/Solr -- there are implementations here of e.g. Punkt, Treebank, and simpler approaches. • nltk.stem: a variety of stemmers using rule-based and data-based algorithms. You'll find familiar ones like Porter and Snowball here
35. On the NLTK menu (2) The following modules are really the core of NLTK: • nltk.grammar: support for context-free grammars (CFGs) which are used in many rule-based systems. Interestingly, CFGs are very much used in computer science theory and programming language design. • nltk.tag: after tokenizing text, you may want to annotate it with metadata that helps with understanding (such as parts of speech). The tag module is solely focused on this task, with classes that help with tagging and retagging tokens, such as Brill and Regexp based taggers. • nltk.chunk: after tagging text, you may find it appropriate to "chunk" the text in order to gain meaning beyond the single-word level. This is particularly handy in information extraction / entity identification.
36. On the NLTK menu (3) Finally, NLTK provides some modules that go beyond actually processing text and onto analyzing large amounts of text for meaning. These include: • nltk.classify: offers feature-based classifiers such as NaiveBayesClassifier and MaxEntClassifier. These are not highly scalable implementaitons, but they are good enough for testing hypotheses and could be made to scale if needed. • nltk.cluster: offers standard algorithms for grouping documents using e.g. the vector space model, k-means, and ways of visualizing these clusters.
37. On the NLTK menu (4) • nltk.collocations: offers simple finders for ngram collocations, e.g. Barack occurs-frequently-with Obama • nltk.featstruct: provides data structures for representing "features" of parsed language constructs. This is often used for "second-pass" filtering of noisy parts of your model. • nltk.corpus.reader.wordnet: a simple wrapper for the powerful Wordnet dictionary/thesaurus.
38. Other options for Python NLP exist • http://www.clips.ua.ac.be/pages/pattern-en ❏ • http://pypi.python.org/pypi/stemming/1.0 ❏ • https://github.com/apresta/tagger ❏ • http://pypi.python.org/pypi/Whoosh/ ❏
39. And other options for NLP generally exist • http://mahout.apache.org/ ❏ • http://incubator.apache.org/opennlp/ ❏ • http://mallet.cs.umass.edu/ ❏ • http://www.cs.waikato.ac.nz/ml/weka/ ❏ • http://alias-i.com/lingpipe/ ❏
40. Brief Interlude for Questions? Next, we dive into doing entity extraction with NLTK. Any questions for now?
41. Entity Extraction What is it, why do we need it?
42. NLTK has a default NER algorithm from nltk import ne_chunk, pos_tag, word_tokenize def entities(text): ne_chunk( pos_tag( word_tokenize(text))) >>> print entities("Steve Jobs created our Apple iPads").pprint() (S (PERSON Steve/NNP) (PERSON Jobs/NNP) created/VBD our/PRP$ shiny/NN (PERSON Apple/NNP iPads/NNP))
43. Good NER is hard So, despite this system's fancy model, including a whole lot of gold-standard data, it still managed to make some mistakes. It considered "Steve" and "Jobs" to be two different people, and it wrongly considered "Apple iPads" to be a person. However, perhaps it is being too ambitious? Can we make it detect "entities" regardless of whether they are geographic regions, people, or other classifications? Yes!
44. Binary NER Binary NER is a simpler problem than "traditional" NER, though still hard. def entities(text): chunks = \ ne_chunk( pos_tag( word_tokenize(text)), binary=True) # binary only enables one type, "NE" return chunks >>> print entities("Steve Jobs created our Apple iPads").pprint() (S (NE Steve/NNP Jobs/NNP) created/VBD our/PRP$ shiny/NN Apple/NNP iPads/NNP)
45. Better, but still not perfect This time, Steve Jobs was properly identified as an entity, but the binary extractor did not pick up on Apple iPad. However, let's think about language a bit. Our trained part-of-speech tagger didn't have a hard time detecting the proper nouns in the sentence. "Apple" and "iPad" were both considered proper nouns, just like "Steve" and "Jobs". For an inclusive NE chunker, wouldn't we be well off to simply treat any proper nouns as entities? We can model this decision with NLTK.
46. RegexpParser for proper nouns from nltk import RegexpParser chunker = RegexpParser(""" NAME: {<NNP>+} """) >>> parsed = chunker.parse(pos_tag(word_tokenize("..."))) >>> print parsed.pprint() (S (NAME Steve/NNP Jobs/NNP) created/VBD our/PRP$ shiny/NN (NAME Apple/NNP iPads/NNP) ) There we go!
47. On the right track Perhaps we can do a combination of traditional NER and syntax rules?
48. Add a small show method to help First, let's make the data easier to inspect. def text2tree(text): chunks = \ ne_chunk( pos_tag( word_tokenize(text)), binary=True) # binary only enables one type, "NE" # I don't normally do this, but it'll help :) def show(self): return self.pprint(margin=40, nodesep=" ->", parens=["", ""]) # MONKEY PATCH chunks.show = types.MethodType(show, chunks) return chunks
49. Trees become entities def chunk2entity(chunk): return ' '.join(leaf[0] for leaf in chunk.leaves()) def tree2entities(tree): # set comprehension, what the fuck up!? entities = { chunk2entity(chunk) for chunk in tree if hasattr(chunk, 'node') } # yea! return entities
50. Even easier to print def p(text): print text2tree(text).show()
51. Unigram problems illustrated (1) First names are valid named entities: >>> p("Angelina just doesn't get Brad") S -> NE -> Angelina/NNP just/RB doesnt/VBZ get/VB NE -> Brad/NNP
52. Unigram problem illustrated (2) Capitalized unigrams lead all sentences: >>> p("Expectation drops in Goldman's earnings") S -> NE -> Expectation/NN drops/NNS in/IN NE -> Goldman/NNP earnings/NNS
53. Unigram problems illustrated (3) Many seeming unigram entities are just things: >>> p("Apple farms reduce output") S -> NE -> Apple/NNP farms/NNS reduce/VB output/NN
54. Contrived examples (1) >>> p("Brad Pitt and Angelina Jolie broken up") S -> NE -> Brad/NNP Pitt/NNP and/CC NE -> Angelina/NNP Jolie/NNP broken/NN up/IN
55. Contrived examples (2) >>> p("Barack Obama gave a speech on the Iraq War") S -> NE -> Barack/NNP Obama/NNP gave/VBD a/DT speech/NN on/IN the/DT NE -> Iraq/NNP War/NNP
56. Contrived examples (3) >>> p("Sachin Kamdar is CEO of Parsely") S -> NE -> Sachin/NNP Kamdar/NNP is/VBZ CEO/NNP of/IN NE -> Parsely/NNP
57. Back to reality Based on these contrived examples, you could draw lots of wrong conclusions. It seems like we're doing a good job, but we're just getting lucky. Headlines (and full text) have a lot more going on than these intentionally simple sentences.
58. Ideas for improvement • Get a better POS tagger • Get a better chunker • Use bigrams, trigrams, or 4-grams • Choose tags to exclude/include • Utilize corpus information (TF/IDF) • Prebuild affinity indices (concordance/collocation) • Leverage a taxonomy (e.g. Wikipedia) • Tap into meta-information (categories) • Navigate up in conceptual understanding (hypernyms) • Use a search engine during NLP phase (fire boolean queries)
59. Single doc vs. corpus analysis
60. Keyword Colocation What is it, why should you care?
61. Colocations Example from nltk.collocations import TrigramCollocationFinder from nltk.metrics import TrigramAssocMeasures from nltk.corpus import webtext from nltk.corpus import stopwords stop_set = set(stopwords.words('english')) stops_filter = lambda w: len(w) < 3 or w in stop_set words = [word.lower() for word in webtext.words('singles.txt')] tcf = TrigramCollocationFinder.from_words(words) tcf.apply_word_filter(stops_filter) tcf.apply_freq_filter(2) tcf.nbest(TrigramAssocMeasures.likelihood_ratio, 4)
62. Stemming What is it, why should you care?
63. Stemming Example from nltk.stem import PorterStemmer from nltk import word_tokenize def publisher_text(pub_name, stem=False): if stem: stemmer = PorterStemmer() stem = stemmer.stem else: stem = lambda word: word textfile = open("data/%s.txt" % pub_name) words = (stem(word.lower()) for headline in textfile for word in word_tokenize(headline.strip().replace(".", ""))) return " ".join(words) publisher_text("bloomberg")
64. Wordnet What is it, why should you care?
65. Wordnet Example >>> from nltk.corpus import wordnet >>> senses = wordnet.synsets("newspaper") >>> s = senses[0] >>> s.hypernym_paths() [[Synset('entity.n.01'), Synset('physical_entity.n.01'), Synset('object.n.01'), Synset('whole.n.02'), Synset('artifact.n.01'), Synset('instrumentality.n.03'), Synset('medium.n.01'), Synset('print_media.n.01'), Synset('press.n.02'), Synset('newspaper.n.01')]] >>> s.hyponyms() [Synset('daily.n.01'), Synset('school_newspaper.n.01'), Synset('tabloid.n.02'), Synset('gazette.n.01')]
66. Classification What is it, why should you care?
67. Classification Example >>> features = [(word, True) for word in word_tokenize(text)] >>> classifier = nltk.data.load('classifiers/my_classifier.pickle') >>> classifier.classify(features) 'pos'
68. How do I make a classifier? NLTK includes a few basic classifiers out of the box. Best option for building your own is japerk's nltk-trainer ❏ project. Provides a CLI for creating classifier pickle files from corpora.
69. nltk-trainer example $ python train_classifier.py --instances paras \ --classifier NaiveBayes \ --ngrams 1 --ngrams 2 \ movie_reviews ... Created in nltk_data/classifiers/movie_reviews_NaiveBayes.pickle
70. Classifier from Scratch from nltk import word_tokenize from nltk.stem.wordnet import WordNetLemmatizer import nltk lemmatizer = WordNetLemmatizer() lemmatize = lambda word: lemmatizer.lemmatize(word.lower()) def get_words(pub): for line in open('data/%s.txt' % pub): for word in word_tokenize(line): yield word def word_features(pub): words = get_words(pub) return {"contains(%s)" % lemmatize(word): True for word in words} bloomberg = [(word_features("bloomberg"), "Bloomberg")] apttherapy = [(word_features("popsugar"), "Apartment Therapy")] feature_set = bloomberg + apttherapy classifier = nltk.NaiveBayesClassifier.train(feature_set)
71. Classifier Results >>> classifier.classify(word_features("apttherapy")) "Apartment Therapy" >>> classifier.classify(word_features("bloomberg")) "Bloomberg" I certainly hope so! But how about others: >>> classifier.classify(word_features("mashable")) "Bloomberg" >>> classifier.classify(word_features("nbclocal")) "Apartment Therapy" Conclusion: • Mashable content is "like" Bloomberg • NBC Local content is "like" Apartment Therapy
72. An obvious improvement We are using a Naive Bayes classifier in above example, but that performs pretty poorly when using single word-based features. (But hey, at least we're lemmatizing.) Better would be to use something like a bigram model or the NER from earlier.
73. Baby Turtles Use your powers wisely, and always remember...
74. Magic Turtles! It's turtles all the way down!
75. Tweet and Meet What did you think? Tweet @amontalenti ❏ with #pydata hash tag! Rate this talk! http://bit.ly/rate-andrew ❏ Connect on LinkedIn: http://linkedin.com/in/andrewmontalenti ❏