Slide 1

Categorizing and Tagging WordsChapter 5 of the NLTK bookPlan for tonightQuizPart of speech taggingUse of the Python dictionary data typeApplication of regular expressionsPlanning for the rest of the semesterUnderstanding textUnderstanding written and spoken text is a very complicated processDistinguishing characteristic of humansMachines do not understandTrying to make machines behave as though they understand has led to interesting insights into the nature of understanding in humansHumans acquire an ability to characterize aspects of language as used in an instance, which aids in seeing meaning in the text or speech.If we want machines to process autonomously, we have to provide explicitly the extra information that humans learn to append to what is written or spoken.

Categorizing and Tagging WordsChapter goals: address these questions:What are lexical categories and how are they used in natural language processing?What is a good Python data structure for storing words and their categories?How can we automatically tag each word of a text with its word class?This necessarily introduces some of the fundamental concepts of natural language processing. Potential use in many application areasWord classesaka Lexical categories.The particular collection of tags used is the tagsetHow nice it would be if there were only one tagset, general enough for all use!Tags for parts of speech identify nouns, verbs, adverbs, adjectives, articles, etcVerbs are further tagged with tense, passive or active voice, etc.There are lots of possibilities for the kinds of tagging desired.First tagging exampleFirst, tokenizeThe tags are cryptic. In this case,CC = coordinating conjunctionRB = adverbIN = prepositionNN = nounJJ = adjective >>> text = nltk.word_tokenize("And now for something completely different")>>> nltk.pos_tag(text)[('And', 'CC'), ('now', 'RB'), ('for', 'IN'), ('something', 'NN'),('completely', 'RB'), ('different', 'JJ')]Exampleimport nltk# from nltk.book import *raw = raw_input("Enter a sentence: ")text = nltk.word_tokenize(raw)print textresult = nltk.pos_tag(text)print result

Enter a sentence: And now for something completely different

['And', 'now', 'for', 'something', 'completely', 'different']

[('And', 'CC'), ('now', 'RB'), ('for', 'IN'), ('something', 'NN'), ('completely', 'RB'), ('different', 'JJ')]HomonymsDifferent words that are spelled the same, but have different meaningsThey may be pronounced differently, or notTags cannot be assigned to the word independently. Word in context required.

More difficult exampleimport nltk# from nltk.book import *raw = raw_input("Enter a sentence: ")text = nltk.word_tokenize(raw)print textresult = nltk.pos_tag(text)print resultEnter a sentence: They refuse to permit us to obtain the refuse permit.['They', 'refuse', 'to', 'permit', 'us', 'to', 'obtain', 'the', 'refuse', 'permit', '.'][('They', 'PRP'), ('refuse', 'VBP'), ('to', 'TO'), ('permit', 'VB'), ('us', 'PRP'), ('to', 'TO'), ('obtain', 'VB'), ('the', 'DT'), ('refuse', 'NN'), ('permit', 'NN'), ('.', '.')]What use is tagging?The goal is autonomous processing that correctly communicates in understandable language.Results: automated telephone treesSpoken directions in gpsDirections provided by mapping programsAny system that uses free text input and provides appropriate information responsesshopping assistants, perhapsMedical diagnosis systemsMany more

Spot checkDescribe a context in which program understanding of free text is needed and/or in which text or spoken response that was not preprogrammed is usefulTagged CorporaIn nltk, tagged token is a tuple (token, tag)This allows us to isolate the two components and use each easilyTags in some corpora are done differently, conversion available >>> tagged_token = nltk.tag.str2tuple('fly/NN')>>> tagged_token('fly', 'NN')>>> tagged_token[0]'fly'>>> tagged_token[1]'NN'Steps to token, tag tuplesOriginal text has each word followed by / and the tag. Change to tokens.Each word and / and tag is a tokenSeparate each token into a word, tag tuple

>>> sent = '''... The/AT grand/JJ jury/NN commented/VBD on/IN a/AT number/NN of/IN... other/AP topics/NNS ,/, among/IN them/PPO the/AT Atlanta/NP and/CC... Fulton/NP-tl County/NN-tl purchasing/VBG departments/NNS which/WDT it/PPS... said/VBD ``/`` zr\\are/BER well/QL operated/VBN and/CC follow/VB generally/RB... accepted/VBN practices/NNS which/WDT inure/VB to/IN the/AT best/JJT... interest/NN of/IN both/ABX governments/NNS ''/'' ./.... '''>>> [nltk.tag.str2tuple(t) for t in sent.split()][('The', 'AT'), ('grand', 'JJ'), ('jury', 'NN'), ('commented', 'VBD'),('on', 'IN'), ('a', 'AT'), ('number', 'NN'), ... ('.', '.')13Handling differing tagging stylesDifferent corpora have different conventions for tagging.NLTK harmonizes those and presents them all as tuplestagged_words() method available for all tagged text in the corpus.Note this will probably not work for arbitrarily selected files. This is done for the files in the corpus. NLTK simplified tagset Simplified Tagset of NLTK

Other languagesNLTK is used for languages other than English, and for alphabets and writing forms other than the Western characters.See the example showing Four Indian Languages (and tell me if they look meaningful!)Looking at text for part of speech>>> from nltk.corpus import brown>>> brown_news_tagged = brown.tagged_words(categories='news', simplify_tags=True)>>> tag_fd = nltk.FreqDist(tag for (word, tag) in brown_news_tagged)>>> tag_fd.keys()['N', 'P', 'DET', 'NP', 'V', 'ADJ', ',', '.', 'CNJ', 'PRO', 'ADV', 'VD', ...]KeyPatterns for nounsnoun tags N for common nouns, NP for proper nounsWhat parts of speech occur before a noun?Construct list of bigramsword-tag pairsFrequency Distribution>>> word_tag_pairs = nltk.bigrams(brown_news_tagged)

>>> list(nltk.FreqDist(a[1] for (a, b) in word_tag_pairs if b[1] == 'N'))

['DET', 'ADJ', 'N', 'P', 'NP', 'NUM', 'V', 'PRO', 'CNJ', '.', ',', 'VG', 'VN', ...]a[0] is the word, a[1] is the pos18A closer looklist(nltk.FreqDist(a[1] for (a, b) in word_tag_pairs if b[1] == 'N'))>>> word_tag_pairs = nltk.bigrams(brown_news_tagged)>>> list(nltk.FreqDist(a[1] for (a, b) in word_tag_pairs if b[1] == 'N'))['DET', 'ADJ', 'N', 'P', 'NP', 'NUM', 'V', 'PRO', 'CNJ', '.', ',', 'VG', 'VN', ...]Lets parse that carefully. What is in word_tag_pairs?

>>> word_tag_pairs[10](("Atlanta's", 'NP'), ('recent', 'ADJ'))

>>> word_tag_pairs[10:15][(("Atlanta's", 'NP'), ('recent', 'ADJ')), (('recent', 'ADJ'), ('primary', 'N')), (('primary', 'N'), ('election', 'N')), (('election', 'N'), ('produced', 'VD')), (('produced', 'VD'), ('``', '``'))]One bigram (Atlantas recent) showing each word with its tag>>> word_tag_pairs[10][1]('recent', 'ADJ')>>> word_tag_pairs[10][1][0]'recentA slice of the word_tag_pairs with red parentheses bracketing each tuple)Patterns for verbsLooking for verbs in the news text and sorting by frequency:>>> wsj = nltk.corpus.treebank.tagged_words(simplify_tags=True)>>> word_tag_fd = nltk.FreqDist(wsj)>>> [word + "/" + tag for (word, tag) in word_tag_fd if tag.startswith('V')]

['is/V', 'said/VD', 'was/VD', 'are/V', 'be/V', 'has/V', 'have/V', 'says/V','were/VD', 'had/VD', 'been/VN', "'s/V", 'do/V', 'say/V', 'make/V', 'did/VD','rose/VD', 'does/V', 'expected/VN', 'buy/V', 'take/V', 'get/V', 'sell/V','help/V', 'added/VD', 'including/VG', 'according/VG', 'made/VN', 'pay/V', ...]for in if -- format for conditional distributionimport nltk# from nltk.book import *wsj=nltk.corpus.treebank.tagged_words(simplify_tags=True)word_tag_fd = nltk.FreqDist(wsj)print "Frequency distribution:"print word_tag_fdverbs= [word+"/"+tag for (word,tag) in word_tag_fd if tag.startswith('V')]print "Verbs: "print verbs[:25]Frequency distribution:

Verbs: ['is/V', 'said/VD', 'are/V', 'was/VD', 'be/V', 'has/V', 'have/V', 'says/V', 'were/VD', 'had/VD', 'been/VN', "'s/V", 'do/V', 'say/V', 'make/V', 'did/VD', 'rose/VD', 'does/V', 'expected/VN', 'buy/V', 'take/V', 'get/V', 'sell/V', 'help/V', 'added/VD']Conditional distributionRecall that conditional distribution requires an event and a condition. We can treat the word as the condition and the tag as the event. >>> cfd1 = nltk.ConditionalFreqDist(wsj)>>> cfd1['yield'].keys()['V', 'N']>>> cfd1['cut'].keys()['V', 'VD', 'N', 'VN']import nltkwsj=nltk.corpus.treebank.tagged_words(simplify_tags=True)cfd1=nltk.ConditionalFreqDist(wsj)word = raw_input("Enter the word to explore: ")if word in cfd1: wordkeys = cfd1[word].keys() print wordkeyselse: print "Word not found"For yield and cut, show the parts of speech.Distributionsor reverse, so the words are the events and we see the tags commonly associated with given words:>>> cfd2 = nltk.ConditionalFreqDist((tag, word) for\ (word, tag) in wsj)>>> cfd2['VN'].keys()['been', 'expected', 'made', 'compared', 'based', 'priced', 'used', 'sold','named', 'designed', 'held', 'fined', 'taken', 'paid', 'traded', 'said', ...]Verb tenseClarifying past tense and past participle, look at some words that are the same for both and the words that are surrounding them: >>> [w for w in cfd1.conditions() if 'VD' in cfd1[w] and 'VN' in cfd1[w]]['Asked', 'accelerated', 'accepted', 'accused', 'acquired', 'added', 'adopted', ...]>>> idx1 = wsj.index(('kicked', 'VD'))>>> wsj[idx1-4:idx1+1][('While', 'P'), ('program', 'N'), ('trades', 'N'), ('swiftly', 'ADV'),('kicked', 'VD')]>>> idx2 = wsj.index(('kicked', 'VN'))>>> wsj[idx2-4:idx2+1][('head', 'N'), ('of', 'P'), ('state', 'N'), ('has', 'V'), ('kicked', 'VN')]Spot checkGet the list of past participles found with cfd2[VN].keys()Collect all the context for each by showing word-tag pair immediately before and immediately after each.>>> cfd2 = nltk.ConditionalFreqDist((tag, word) for\ (word, tag) in wsj)Adjectives and Adverbs and other parts of speech. We can find them all, examine their context, etc.Example, looking at a wordoften how is it used in common English usage?>>> brown_learned_text = brown.words(categories='learned')>>> sorted(set(b for (a, b) in nltk.ibigrams(brown_learned_text) if a == 'often'))[',', '.', 'accomplished', 'analytically', 'appear', 'apt', 'associated', 'assuming','became', 'become', 'been', 'began', 'call', 'called', 'carefully', 'chose', ...]This gives us a set of sorted words, from the bigrams in the indicated text, where the first word is often>>> brown_lrnd_tagged = brown.tagged_words(categories='learned', simplify_tags=True)>>> tags = [b[1] for (a, b) in nltk.ibigrams(brown_lrnd_tagged) if a[0] == 'often']>>> fd = nltk.FreqDist(tags)>>> fd.tabulate() VN V VD DET ADJ ADV P CNJ , TO VG WH VBZ . 15 12 8 5 5 4 4 3 3 1 1 1 1 1This gives us the distribution of the part of speech of the word following often in sorted orderFinding patternsWe saw how to use regular expressions to extract patterns on words, now lets extract patterns of parts of speech.We look at each three-word phrase in the text, and find to :from nltk.corpus import browndef process(sentence): for (w1,t1), (w2,t2), (w3,t3) in \ nltk.trigrams(sentence): if (t1.startswith('V') and t2 == 'TO' and \ t3.startswith('V')): print w1, w2, w3 >>> for tagged_sent in brown.tagged_sents():... process(tagged_sent)...combined to achievecontinue to placeserve to protectwanted to waitallowed to placeexpected to become...POS ambiguitiesLooking at how the words are tagged, may help understand the tagging>>> brown_news_tagged = brown.tagged_words(categories='news',\ simplify_tags=True)>>> data = nltk.ConditionalFreqDist((word.lower(), tag)... for (word, tag) in brown_news_tagged)>>> for word in data.conditions():... if len(data[word]) > 3:... tags = data[word].keys()... print word, ' '.join(tags)...best ADJ ADV NP Vbetter ADJ ADV V DETclose ADV ADJ V Ncut V N VN VDeven ADV DET ADJ VRecall that a Conditional Frequency Distribution (ConditionalFreqDist()has an event and a condition.Each element of data has data.conditions() and each condition has keysdata[].keys()Python DictionaryAllows mapping between arbitrary types (not necessary to have a numeric index)Note the addition of defaultdictreturns a value for a non-existing entryuses the default value for the type0 for number, [] for empty list, etc.We can specify a default value to useTagging rare words with default value>>> alice = nltk.corpus.gutenberg.words('carroll-alice.txt')>>> vocab = nltk.FreqDist(alice)>>> v1000 = list(vocab)[:1000]>>> mapping = nltk.defaultdict(lambda: 'UNK')>>> for v in v1000:... mapping[v] = v...>>> alice2 = [mapping[v] for v in alice]>>> alice2[:100]['UNK', 'Alice', "'", 's', 'Adventures', 'in', 'Wonderland', 'by', 'UNK', 'UNK','UNK', 'UNK', 'CHAPTER', 'I', '.', 'UNK', 'the', 'Rabbit', '-', 'UNK', 'Alice','was', 'beginning', 'to', 'get', 'very', 'tired', 'of', 'sitting', 'by', 'her',A lambda function is an unnamed function that can be defined and used wherever a function object is required.Spot checkLook at the previous exampleFor each line of the code, determine exactly what it does. Think first, then insert print commands to show each result.Incrementally updating a dictionaryInitialize an empty defaultdictProcess each part of speech tag in the textIf it has not been seen before, it will have zero countEach time the tag is seen, increment its counter

itemgetter(n) returns a function that can be called on some other sequence object to obtain the nth element

>>> counts = nltk.defaultdict(int)>>> from nltk.corpus import brown>>> for (word, tag) in brown.tagged_words(categories='news'):... counts[tag] += 1...>>> counts['N']22226>>> list(counts)['FW', 'DET', 'WH', "''", 'VBZ', 'VB+PPO', "'", ')', 'ADJ', 'PRO', '*', '-', ...]

>>> from operator import itemgetter>>> sorted(counts.items(), key=itemgetter(1), reverse=True)[('N', 22226), ('P', 10845), ('DET', 10648), ('NP', 8336), ('V', 7313), ...]>>> [t for t, c in sorted(counts.items(), key=itemgetter(1), reverse=True)]['N', 'P', 'DET', 'NP', 'V', 'ADJ', ',', '.', 'CNJ', 'PRO', 'ADV', 'VD', ...]Look at the parameters of sorted. What does each represent?What to sortWhat is the sort keyWhat orderAnother pattern for updatingimport nltklast_letters = nltk.defaultdict(list)words = nltk.corpus.words.words('en')for word in words: key = word[-2:] last_letters[key].append(word)print last_letters['ly']

['abactinally', 'abandonedly', 'abasedly', 'abashedly', 'abashlessly', 'abbreviately','abdominally', 'abhorrently', 'abidingly', 'abiogenetically', 'abiologically', ...]Note that each entry in the dictionary has a unique keyThe value part of the entry is a listindexBecause accumulating a list of words is so common, NLTK defines a defaultdict(list) nltk.Index >>> anagrams = nltk.Index((''.join(sorted(w)), w) for w in words)>>> anagrams['aeilnrt']['entrail', 'latrine', 'ratline', 'reliant', 'retinal', 'trenail']What does this do? >>> anagrams = nltk.defaultdict(list)>>> for word in words:... key = ''.join(sorted(word))... anagrams[key].append(word)...>>> anagrams['aeilnrt']Exactly the same as this:Inverted dictionary >>> pos.update({'cats': 'N', 'scratch': 'V', 'peacefully': 'ADV', 'old': 'ADJ'})>>> pos2 = nltk.defaultdict(list)>>> for key, value in pos.items():... pos2[value].append(key)...>>> pos2['ADV']['peacefully', 'furiously']>>> pos = {'colorless': 'ADJ', 'ideas': 'N', 'sleep': 'V', 'furiously': 'ADV'}>>> pos2 = dict((value, key) for (key, value) in pos.items())>>> pos2['N']'ideas'pos is a dictionary by { }pos2 is a dictionary by dict( )>>> pos{'furiously': 'ADV', 'sleep': 'V', 'ideas': 'N', 'colorless': 'ADJ'}>>> pos2{'ADV': 'furiously', 'N': 'ideas', 'ADJ': 'colorless', 'V': 'sleep'}Automatic taggingDefaultTag each token with the most common tag.>>> tags = [tag for (word, tag) in brown.tagged_words(categories='news')]>>> nltk.FreqDist(tags).max()'NN'>>> raw = 'I do not like green eggs and ham, I do not like them Sam I am!'>>> tokens = nltk.word_tokenize(raw)>>> default_tagger = nltk.DefaultTagger('NN')>>> default_tagger.tag(tokens)[('I', 'NN'), ('do', 'NN'), ('not', 'NN'), ('like', 'NN'), ('green', 'NN'),('eggs', 'NN'), ('and', 'NN'), ('ham', 'NN'), (',', 'NN'), ('I', 'NN'),('do', 'NN'), ('not', 'NN'), ('like', 'NN'), ('them', 'NN'), ('Sam', 'NN'),('I', 'NN'), ('am', 'NN'), ('!', 'NN')]>>> default_tagger.evaluate(brown_tagged_sents)0.13089484257215028Not very good!NN (noun) is the most common tag in the given text.NLTK includes an evaluate function for each taggerRegular expression taggerUse expected patterns to assign tags>>> patterns = [... (r'.*ing$', 'VBG),# gerunds... (r'.*ed$', 'VBD'), # simple past... (r'.*es$', 'VBZ'), # 3rd singular present... (r'.*ould$', 'MD'), # modals... (r'.*\'s$', 'NN$'), # possessive nouns... (r'.*s$', 'NNS'), # plural nouns... (r'^-?[0-9]+(.[0-9]+)?$', 'CD'), # cardinal num.... (r'.*', 'NN') # nouns (default)... ] >>> regexp_tagger = nltk.RegexpTagger(patterns)>>> regexp_tagger.tag(brown_sents[3])[('``', 'NN'), ('Only', 'NN'), ('a', 'NN'), ('relative', 'NN'), ('handful', 'NN'),('of', 'NN'), ('such', 'NN'), ('reports', 'NNS'), ('was', 'NNS'), ('received', 'VBD'),("''", 'NN'), (',', 'NN'), ('the', 'NN'), ('jury', 'NN'), ('said', 'NN'), (',', 'NN'),('``', 'NN'), ('considering', 'VBG'), ('the', 'NN'), ('widespread', 'NN'), ...]>>> regexp_tagger.evaluate(brown_tagged_sents)0.20326391789486245betterTag applied as soon as a match is found. If no match, defaults to NNLookup taggerFind the most common words and store their usual tag. >>> fd = nltk.FreqDist(brown.words(categories='news'))>>> cfd = nltk.ConditionalFreqDist(brown.tagged_words(categories='news'))>>> most_freq_words = fd.keys()[:100]>>> likely_tags = dict((word, cfd[word].max()) for word in most_freq_words)>>> baseline_tagger = nltk.UnigramTagger(model=likely_tags)>>> baseline_tagger.evaluate(brown_tagged_sents)0.45578495136941344Still betterLook at this carefullyRefined lookupAssign tags to words that are not nouns, and default others to noun. >>> baseline_tagger = nltk.UnigramTagger(model=likely_tags,... backoff=nltk.DefaultTagger('NN'))Model size and performance

EvaluationGold standard test dataCorpus that has been manually annotated and carefully evaluated. Test the tagging technique against the test case, where the right answers are known. If it does well there, assume it does well in general.N-gram taggingUnigramUse the most frequent tag for a wordMust have a gold standard for reference>>> from nltk.corpus import brown>>> brown_tagged_sents = brown.tagged_sents(categories='news')>>> brown_sents = brown.sents(categories='news')

>>> unigram_tagger = nltk.UnigramTagger(brown_tagged_sents)>>> unigram_tagger.tag(brown_sents[2007])

[('Various', 'JJ'), ('of', 'IN'), ('the', 'AT'), ('apartments', 'NNS'),('are', 'BER'), ('of', 'IN'), ('the', 'AT'), ('terrace', 'NN'), ('type', 'NN'),(',', ','), ('being', 'BEG'), ('on', 'IN'), ('the', 'AT'), ('ground', 'NN'),('floor', 'NN'), ('so', 'QL'), ('that', 'CS'), ('entrance', 'NN'), ('is', 'BEZ'),('direct', 'JJ'), ('.', '.')]>>> unigram_tagger.evaluate(brown_tagged_sents)0.9349006503968017Testing on the same data as training. Separate training and testing >>> size = int(len(brown_tagged_sents) * 0.9)>>> size4160>>> train_sents = brown_tagged_sents[:size]>>> test_sents = brown_tagged_sents[size:]>>> unigram_tagger = nltk.UnigramTagger(train_sents)>>> unigram_tagger.evaluate(test_sents)0.81202033290142528The testing data is now different from the training data. So, this is a better test of the process.If training and evaluation are on the same data, we certainly expect a very good performance! More realistically, train on part of the data and test on the rest.train on the last 4160 sentencestest on the rest of the sentencesYour turnExperiment with this tagger.Does it matter if you train on the first part or the last part?What is the effect of training on 80% of the data and testing on the other 20%Notice that the training and testing, though on different sentences, is all from the same category of the Brown corpus. How well would you expect the training to do in a different corpus? If the corpus was also from a news category? If it was from a novel?General N-Gram TaggingCombine current word and the part of speech tags of the previous n-1 words to give the current word some context.

Bigram tagger >>> bigram_tagger = nltk.BigramTagger(train_sents)>>> bigram_tagger.tag(brown_sents[2007])

[('Various', 'JJ'), ('of', 'IN'), ('the', 'AT'), ('apartments', 'NNS'),('are', 'BER'), ('of', 'IN'), ('the', 'AT'), ('terrace', 'NN'),('type', 'NN'), (',', ','), ('being', 'BEG'), ('on', 'IN'), ('the', 'AT'),('ground', 'NN'), ('floor', 'NN'), ('so', 'CS'), ('that', 'CS'),('entrance', 'NN'), ('is', 'BEZ'), ('direct', 'JJ'), ('.', '.')]

>>> unseen_sent = brown_sents[4203]>>> bigram_tagger.tag(unseen_sent)[('The', 'AT'), ('population', 'NN'), ('of', 'IN'), ('the', 'AT'), ('Congo', 'NP'),('is', 'BEZ'), ('13.5', None), ('million', None), (',', None), ('divided', None),('into', None), ('at', None), ('least', None), ('seven', None), ('major', None),('``', None), ('culture', None), ('clusters', None), ("''", None), ('and', None),('innumerable', None), ('tribes', None), ('speaking', None), ('400', None),('separate', None), ('dialects', None), ('.', None)] >>> bigram_tagger.evaluate(test_sents)0.10276088906608193Reliance on context not seen in training reduces accuracy.This is the precision recall tradeoff of information retrievalSo, whats the problem?Now, we are matching pairs of words and a pair is less likely to have occurred before. Using context provides greater accuracy, when it is able to find a match. However, frequently, it will not find a match.Compromise use the most accurate tagger for what it can do, then back it up with another tagger for the parts that do not get tagged.Combining taggersUse the benefits of several types of taggersTry the bigram taggerWhen it is unable to find a tag, use the unigram taggerIf that fails, then use the default tagger >>> t0 = nltk.DefaultTagger('NN')>>> t1 = nltk.UnigramTagger(train_sents, backoff=t0)>>> t2 = nltk.BigramTagger(train_sents, backoff=t1)>>> t2.evaluate(test_sents)0.84491179108940495Spot CheckNote the order of appearance of the taggers. Why is that?Extend to a trigramtagger, t3, which backs off to t2.Assignment for Chapter 5Assignment for two weeks:NLTK chapter 5: # 3, 12 Program: (be sure to work alone!) Choose one of these: 5.25, 5.30, 5.33