Khaled Notes

02 Pre-processing - Chunk 02 (s.31-60)

Properties1
tagsmidterm, text-web-mining

20 min read

02 Pre-processing - Chunk 02 (s.31-60)

Overview

This part of the Pre-processing lecture finishes the normalisation topics (Concept - Lemmatization, Concept - Morphology, Concept - Stemming, with Porter vs Lancaster stemmers in NLTK) (s.31-35) and then covers sentence boundary detection, i.e. deciding which periods end a sentence, using a hand-written decision-tree rule set (s.36-41). It then moves to syntactic analysis: Concept - POS Tagging with the Penn Tree Bank tag set (s.42-48), Concept - Word Sense Disambiguation with the (Simplified) Concept - Lesk Algorithm and Babelfy (s.49-52), and Concept - Parsing with phrase-structure trees and Concept - Universal Dependencies (s.53-56). The chunk closes by noting that off-the-shelf tools are used in about 80% of applications (s.57) and by introducing Concept - WordNet (s.58-60).

Detailed notes

1. Lemmatization (s.31)

  • Additional examples (s.31): lemmatization reduces variant forms to the base form.
    • am, are, is -> be
    • car, cars, car’s, cars’ -> car
    • the boy’s cars are different colors -> the boy car be different color
  • Lemmatization has to find the correct dictionary headword form (i.e. the root or lemma form) (s.31).
  • Example tool: Stanford CoreNLP supports lemmatization (http://stanfordnlp.github.io/CoreNLP/, online demo http://nlp.stanford.edu:8080/corenlp/) (s.31).
  • Box on the slide, “Stanford CoreNLP Lemmatization”: denormalization -> denormalization; reapplications -> reapplication; reapplied -> reapply (s.31). (Takeaway: the lemma is a valid word; plural -s and past -ied are normalised, but derivational prefixes/suffixes such as de-, re-, -ization, -ation are kept.)

2. Morphology: morphemes, stems, affixes (s.32)

  • Morphemes: the small meaningful units that make up words. E.g. un-like-ly contains three morphemes (s.32).
  • Stems: the main part of a word that stays the same when endings are added. E.g. writ is the stem of writes, writing, written (s.32).
  • Affixes: bits and pieces that adhere to stems (i.e. the prefix and suffix); often with grammatical functions. E.g. likes (the -s) (s.32).
  • See Concept - Morphology.

3. Stemming (s.33-35)

  • Inflectional stemming: when normalisation is confined to regularizing grammatical variants such as singular/plural and present/past (s.33). This is called “morphological analysis” (s.33).
  • English has many irregular word forms and non-intuitive spelling, so it is more difficult, e.g. sought -> seek (s.33).
  • In English, an inflectional-stemming algorithm must be part rule-based and part dictionary-based (s.33).
  • Any English stemming algorithm that operates only on tokens, without more grammatical information such as part-of-speech, will make some mistakes because of ambiguity (s.33). Example: is “bored” the adjective (“he is bored”) or the past tense of the verb “bore”? (“He bored her with his stories about military life.”) (s.33)
  • Stemming in general (s.34):
    • Reduces terms to their stems; used in information retrieval and text mining applications.
    • Stemming is crude chopping of affixes; it is language dependent.
    • e.g. automates, automatic, automation are all reduced to automat.
  • Python NLTK demo (s.35): see Figures. Porter and Lancaster stemmers give different outputs for the same sentence.
InputPorter outputLancaster output
”Stemming is funnier than a bummer says the sushi loving computer scientist”Stem is funnier than a bummer say the sushi love comput scientiststem is funny than a bum say the sush lov comput sci

(s.35) Takeaway: Lancaster is more aggressive (sushi -> sush, loving -> lov, scientist -> sci, funnier -> funny, bummer -> bum) than Porter.

4. Sentence boundary detection (s.36-41)

  • s.36 is a section-divider slide (“Sentence Boundary Detection”).
  • Why needed (s.37): more sophisticated linguistic parsing often requires a complete sentence as input, e.g. sentence-level sentiment analysis. Other information extraction algorithms also operate on one sentence at a time.
  • The problem (s.37): decide which instances of a period (.) followed by whitespace are sentence delimiters and which are not, because ? and ! are assumed to be unambiguous sentence boundaries.
  • Sentence segmentation (s.38):
    • !, ? are relatively unambiguous.
    • Period ”.” is quite ambiguous: it can be a sentence boundary, part of an abbreviation (Inc., Dr.), or part of a number (.02%, 4.3).
    • Build a binary classifier that looks at a ”.”, and decides EndOfSentence / NotEndOfSentence. Classifiers can be hand-written rules, regular expressions, or machine learning.
  • Decision tree for EOS (s.39), questions asked in order:
    1. “Lots of blank lines after me?” YES -> E-O-S; NO -> next.
    2. “Final punctuation is ?, !, or :?” YES -> E-O-S; NO -> next.
    3. “Final punctuation is period” YES -> next; NO -> Not E-O-S.
    4. “I am ‘etc’ or other abbreviation” YES -> Not E-O-S; NO -> E-O-S.
  • Implementing decision trees (s.40):
    • A decision tree is just an if-then-else statement; we can think of the questions in the tree.
    • The interesting research is choosing the features.
    • Setting up the structure is often too hard by hand; hand-building is only possible for very simple features and domains. For numeric features it is too hard to pick each threshold.
    • Instead, the structure is usually learned by machine learning from a training corpus.
    • The features can be exploited by any kind of classifier: SVM, Neural Networks, Logistic Regression, etc.
  • Hand-written rule algorithm (s.41), see Concept - Sentence Boundary Detection:
    • Input: a text with periods. Output: the same text with End-of-Sentence (EOS) periods identified.
    • Overall strategy: (1) replace all identifiable non-EOS periods with another character; (2) apply rules to all remaining periods in the text and mark EOS periods; (3) retransform the characters from step 1 back to non-EOS periods; (4) the text now has all EOS periods clearly identified.
    • Rules: all ? ! are EOS; if ” or ’ appears before the period, it is EOS; if the following character is not white space, it is not EOS; if ) } ] appears before the period, it is EOS; if the token to which the period is attached is capitalised and is < 5 characters and the next token begins uppercase, it is not EOS; if the token has other periods, it is not EOS; if the token begins with a lowercase letter and the next token after whitespace is uppercase, it is EOS; if the token has < 2 characters, it is not EOS; if the next token after whitespace begins with $ ( { [ ” ’ it is EOS; otherwise, the period is not EOS.
    • Examples of EOS: period after a closing quote/bracket, e.g. ...". ...'. ...). ...}. ...]. and ...x.Y; period followed by $ ( { [ ” ’ (opening symbols).
    • Examples of not EOS: Ph.D., www.google.com, i.e.

5. Syntactic analysis overview (s.42-43)

  • s.42 is a section-divider slide; s.43 lists the three topics: Part-of-Speech Tagging, Word Sense Disambiguation, Parsing.

6. Part-of-Speech (POS) tagging (s.44-48)

  • If no further linguistic analysis is necessary one can go directly to feature generation using the tokens (e.g. linguistic and analysis from the sentence) (s.44).
  • If the goal is more specific, e.g. recognising names of people, places and organisations, it is usually desirable to perform additional linguistic analyses and extract more sophisticated features, e.g. San Francisco (s.44).
  • In English some analyses use as few as six or seven categories, others nearly one hundred; most English grammars have at minimum noun, verb, adjective, adverb, preposition, conjunction (s.44).
  • Uses of POS (s.45): feature reduction, e.g. use only verb, adjective and adverb for sentiment classification; the distribution of POS can be used for author, gender and document genre (formal vs informal) classification.
  • Resources (s.46):
    • Penn Tree Bank uses a set of 36 categories, constructed from the Wall Street Journal corpus. A tree bank is a parsed text corpus that annotates sentence structure such as POS and phrases. See Concept - Penn Treebank.
    • Almost all POS taggers are trained on the Wall Street Journal corpus available from the LDC (Linguistic Data Consortium). E.g. I love you -> I (personal pronoun) love (verb, not noun).
    • The Brill tagger is in the public domain and widely used (online version available).
    • The Stanford Parser is a statistical parser (Java implementation).
  • Penn Tree Bank POS set (s.47), excerpt of the table shown (tag - description):
TagDescriptionTagDescription
CCCoordinating conjunctionNNSNoun, plural
CDCardinal numberPOSPossessive ending
DTDeterminerUHInterjection
EXExistential thereVBVerb, base form
FWForeign wordVBDVerb, past tense
INPreposition or subordinating conjunctionVBGVerb, gerund or present participle
JJAdjectiveVBNVerb, past participle
JJRAdjective, comparativeVBPVerb, non-3rd person singular present
JJSAdjective, superlativeWDTWh-determiner
LSList item markerMDModal
NNNoun, singular or mass

The slide notes that the full list of POS categories is at a linked NYU page (the table shown is only a part of the 36).

  • Stanford Parser online demo (s.48): input “My dog also likes eating sausage.” gives the tagging My/PRP$ dog/NN also/RB likes/VBZ eating/VBG sausage/NN ./. Note box: PRP$ = possessive pronoun, RB = adverb, VBZ = verb, 3rd person singular present.

7. Word Sense Disambiguation (WSD) (s.49-52)

  • Task (s.49): disambiguate “bank” in “The bank can guarantee deposits will eventually cover future tuition costs because it invests in adjustable-rate mortgage securities.” given two WordNet senses:
    • bank^1 - Gloss: a financial institution that accepts deposits and channels the money into lending activities. Examples: “he cashed a check at the bank”, “that bank holds the mortgage on my home”.
    • bank^2 - Gloss: sloping land (especially the slope beside a body of water). Examples: “they pulled the canoe up on the bank”, “he sat on the bank of the river and watched the currents”.
  • Simplified Lesk algorithm (s.50): choose the sense with the most word overlap between the gloss (and examples) and the context, not counting stop words. On the slide, the words deposits and mortgage in the sentence (boxed in green) overlap with bank^1 (gloss “deposits”, example “mortgage”), so bank^1 wins; bank^2 has no overlap. See Concept - Lesk Algorithm.
  • NLTK Lesk (s.51): performs the classic Lesk algorithm; given an ambiguous word and the context, Lesk returns the Synset with the highest number of overlapping words between the context sentence and the definitions of each Synset. Screenshot: lesk(['I','went','to','the','bank','to','deposit','money','.'], 'bank', 'n') returns Synset('savings_bank.n.02') (same without the POS argument); a second screenshot lists the WordNet synsets of ‘bank’ with definitions (bank.n.01 sloping land; depository_financial_institution.n.01 a financial institution that accepts deposits…; bank.n.03 a long ridge or pile; bank.n.04 an arrangement of similar objects in a row or in tiers; bank.n.05 a supply or stock held in reserve; bank.n.06 the funds held by a gambling house; savings_bank.n.02 a container (usually with a slot in the top) for keeping money at home; bank.n.07 a slope in the turn of a road; bank.n.09 a building in which the business of banking is transacted; bank.n.10 a flight maneuver; and verb senses bank.v.01-v.07 such as tip laterally, enclose with a bank, do business with a bank, act as the banker in a game, be in the banking business, deposit.v.02 put into a bank account, cover with ashes, trust.v.01 have confidence). Takeaway: “bank” has many WordNet senses (nouns and verbs), so WSD is needed.
  • Babelfy (s.52): state-of-the-art system based on the BabelNet multilingual semantic network, for multilingual WSD and Entity Linking (babelfy.org). Screenshot: input “I went to the bank to deposit my salary”, language buttons (English, Arabic, Chinese, French, German, Greek, Hebrew, Hindi, Italian, Japanese, Russian), output highlights bank, deposit, salary as concepts (green) with glosses (bank: “A financial institution that accepts deposits and channels the money into lending activities”; deposit: “Put into a bank account”; salary: “a form of periodic payment from an employer to an employee”).

8. Parsing (s.53-56)

  • Parsing is the step of producing a full parse of a sentence; each word is connected to a single structure, usually a tree (s.53).
  • Considerable research has built parsers from a statistical analysis of tree banks of sentences parsed by hand (s.53).
  • Why such an expensive process? It provides detailed syntactic relationship information that phrase identification cannot provide (s.53).
  • Phrase structure tree (s.53 and s.54): the tree for “the dog saw a man in the park” (S -> NP VP; VP -> V NP; NP -> Det N PP; PP -> P NP) on s.53; for “Johnson was replaced at XYZ Corp. by Smith” (S -> NP(N Johnson), VP -> [VP(AUX was, PPART replaced), PP(PREP at, PNOUN XYZ Corp), PP(PREP by, PNOUN Smith)]) on s.54 (Fig. 2.9 Simple parse tree).
    • Takeaway s.54: by looking at the parse tree a machine can infer that Johnson was replaced by Smith (correct) and not that Steve was replaced by Smith (wrong), for the more complex sentence “Johnson who is son of Steve was replaced at XYZ Corp. by Smith”; the bracketed Stanford output (ROOT (S (NP (NP (NNP Johnson)) (SBAR (WHNP (WP who)) (S (VP (VBZ is) (NP (NP (NN son)) (PP (IN of) (NP (NNP Steve)))))))) (VP (VBD was) (VP (VBN replaced) (PP (IN at) (NP (NNP XYZ) (NNP Corp.))) (PP (IN by) (NP (NNP Smith))))) (. .))) is shown.
  • Universal dependencies (s.55), i.e. grammatical relations, evolved out of Stanford Dependencies. For “I like him”: tree with root like, edge nsubj to I and dobj to him. Stanford Parser output: nsubj(like-2, I-1), root(ROOT-0, like-2), dobj(like-2, him-3). Stanford CoreNLP “Basic Dependencies” display: I (PRP) <-nsubj- like (VBP) -dobj-> him (PRP).
  • s.56, “Johnson was replaced at XYZ Corp. by Smith” (CoreNLP): nsubjpass(replaced-3, Johnson-1), auxpass(replaced-3, was-2), root(ROOT-0, replaced-3), case(Corp.-6, at-4), compound(Corp.-6, XYZ-5), nmod(replaced-3, Corp.-6), case(Smith-8, by-7), nmod(replaced-3, Smith-8). Tags: Johnson NNP, was VBD, replaced VBN, at IN, XYZ NNP, Corp. NNP, by IN, Smith NNP. See Concept - Universal Dependencies.

9. Off-the-shelf tools (s.57)

  • For tasks such as tokenization, POS and parsing, we can build machine learning models from scratch, but in almost 80% of applications the off-the-shelf tools are used: NLTK, Stanford CoreNLP, Spacy and Textblob (s.57).

10. WordNet (s.58-60)

  • s.58 is a section-divider slide (“WordNet: Linguistic Resources”).
  • WordNet is the most well developed and widely used lexical database for English (s.59). It consists of 4 databases: nouns, verbs, adjectives and adverbs (s.59). Online version at wordnetweb.princeton.edu.
  • Each database consists of sense entries, each being a set of synonyms (synsets), e.g. {musician, instrumentalist, player}; {person, individual, someone}; {life form, organism, being} (s.59). Screenshot: WordNet Search 3.1 for “musician”, showing noun senses “S: (n) musician, instrumentalist, player (someone who plays a musical instrument (as a profession))” and “S: (n) musician (artist who composes or conducts music as a profession)”; key: “S:” = show synset (semantic) relations, “W:” = show word (lexical) relations.
  • Size of WordNet (s.60):
CategoryUnique formsNumber of senses
Noun94474116317
Verb1031922066
Adjective2017029881
Adverb45465677

(Takeaway: nouns dominate; more senses than unique forms in every category = polysemy.) See Concept - WordNet.

Figures & diagrams

  • s.31 Red box “Stanford CoreNLP Lemmatization” with three examples (denormalization, reapplications -> reapplication, reapplied -> reapply).
  • s.32 Slide text colour-highlights the stem writ in writes/writing/written and the affix s in likes (affix/stem distinction).
  • s.34 Two red boxes with an arrow: “for example compressed and compression are both accepted as equivalent to compress.” -> “for exampl compress and compress ar both accept as equival to compress” (Porter-style stemming; stems such as exampl, equival, ar are not real words).
  • s.35 Two screenshots of an online NLTK “Stem Text” tool: Porter vs Lancaster output (table in section 3).
  • s.39 Decision tree for end-of-sentence classification (blue question nodes, E-O-S leaves in blue, Not E-O-S leaves in red).
  • s.41 Left: lists of EOS / not-EOS examples; right: boxed pseudo-code “End-of-sentence detection algorithm” (input/output, 4-step strategy, rules).
  • s.47 Table of the Penn Tree Bank POS tags (CC … WDT), with link to the full list.
  • s.48 Stanford Parser web screenshot with tagged sentence and note box.
  • s.49-50 Table of the two WordNet senses of bank; s.50 adds green boxes marking overlapping words deposits and mortgage.
  • s.51 NLTK Python console: lesk calls and the list of WordNet synsets/definitions of bank.
  • s.52 Babelfy screenshot: sentence “I went to the bank to deposit my salary” with highlighted concepts and gloss cards.
  • s.53 Parse tree of “the dog saw a man in the park”.
  • s.54 Phrase structure tree (Fig. 2.9) plus a bracketed parse of the longer sentence.
  • s.55-56 Dependency graphs and typed-dependency lists (nsubj, dobj, nsubjpass, auxpass, case, compound, nmod, root).
  • s.59 WordNet web search screenshot; s.60 table of unique forms and senses per category.
flowchart TD
    A[Period '.' in text] --> B{Lots of blank lines after me?}
    B -- YES --> E1[E-O-S]
    B -- NO --> C{Final punctuation is ? ! or :}
    C -- YES --> E2[E-O-S]
    C -- NO --> D{Final punctuation is period?}
    D -- NO --> N1[Not E-O-S]
    D -- YES --> F{I am 'etc' or other abbreviation?}
    F -- YES --> N2[Not E-O-S]
    F -- NO --> E3[E-O-S]

flowchart LR
    SA[Syntactic Analysis] --> POS[POS Tagging<br/>Penn Tree Bank 36 tags]
    SA --> WSD[Word Sense Disambiguation<br/>Lesk, Babelfy, WordNet]
    SA --> P[Parsing<br/>phrase structure tree<br/>universal dependencies]
    WN[WordNet<br/>synsets] --> WSD
    TB[Tree bank<br/>WSJ corpus] --> POS
    TB --> P

flowchart TD
    S[Sentence: bank ... deposits ... mortgage] --> L[Simplified Lesk: count overlap<br/>gloss + examples vs context, ignore stop words]
    L --> S1[bank1 financial institution<br/>overlap: deposits, mortgage = 2]
    L --> S2[bank2 sloping land<br/>overlap = 0]
    S1 --> R[Choose bank1]

Key terms

Formulas

No formulas appear on these slides (s.31-60). The only quantitative item is the Lesk overlap score: sense = argmax over senses of excluding stop words (s.50-51; notation added, slide describes it in words).

Likely exam questions

Style follows Past Exam - Question Bank & Style Guide (short “discuss / describe / explain” answers, compare tables, algorithm steps).

  1. Compare stemming and lemmatization, with examples. (s.31, s.33-34)
StemmingLemmatization
What it doesCrude chopping of affixes (s.34)Finds the correct dictionary headword / lemma (s.31)
Exampleautomates, automatic, automation -> automatam, are, is -> be; cars -> car
Outputmay not be a word (exampl, equival)valid dictionary form
Needsrules (language dependent)dictionary (+ POS)
  1. Define morpheme, stem and affix. (s.32) Morphemes: smallest meaningful units (un-like-ly = 3). Stem: main part that stays the same (writ in writes/writing/written). Affixes: prefixes/suffixes adhering to stems, often with grammatical function (likes).
  2. What is inflectional stemming and why is English hard? (s.33) Normalisation restricted to grammatical variants (singular/plural, present/past) = morphological analysis. English has irregular forms and non-intuitive spelling (sought -> seek), so the algorithm must be part rule-based, part dictionary-based, and token-only stemmers err due to ambiguity (bored: adjective vs past tense of bore).
  3. Compare the Porter and Lancaster stemmers. (s.35) Same input gives different stems; Lancaster is more aggressive (sushi -> sush, loving -> lov, scientist -> sci) than Porter (sushi, love, scientist -> comput scientist).
  4. Why do we need sentence boundary detection and why is it difficult? (s.37-38) Parsers/sentiment/IE need whole sentences; ? and ! are unambiguous but ”.” may mark sentence end, abbreviations (Inc., Dr.) or numbers (.02%, 4.3).
  5. Describe the hand-written end-of-sentence algorithm. (s.41) Replace identifiable non-EOS periods with another character; apply rules to mark EOS periods; retransform step-1 characters back; rules include: all ? ! are EOS; following char not whitespace -> not EOS; token with other periods -> not EOS; abbreviation (<2 chars, capitalised and <5 chars followed by uppercase) -> not EOS; etc.
  6. Draw the decision tree for EOS. (s.39) Blank lines -> EOS; final punctuation ? ! : -> EOS; period -> abbreviation? yes: not EOS, no: EOS; otherwise not EOS.
  7. Why are decision trees usually learned rather than hand-built? (s.40) Hand-building only works for very simple features/domains; numeric thresholds are too hard to pick; structure is learned from a training corpus; features can feed any classifier (SVM, NN, logistic regression).
  8. What is POS tagging and what is it used for? (s.44-46) Assigning a grammatical category to each token; used for feature reduction (verbs/adjectives/adverbs for sentiment), entity recognition, author/gender/genre classification via POS distribution. Penn Tree Bank: 36 tags, trained on WSJ corpus.
  9. What is a tree bank? (s.46) A parsed text corpus annotating sentence structure such as POS and phrases.
  10. Explain word sense disambiguation using the Simplified Lesk algorithm on “bank”. (s.49-50) Choose the sense whose gloss/examples overlap most with the context (ignoring stop words): deposits and mortgage overlap with bank^1 (financial institution), so bank^1 is chosen over bank^2 (sloping land).
  11. What is parsing and why is it worth its cost? (s.53) Full parse of a sentence connecting every word into one structure (usually a tree); gives detailed syntactic relationships that phrase identification cannot, e.g. who was replaced by whom (s.54).
  12. What is a universal dependency? Give an example. (s.55-56) Grammatical relation between words, evolved from Stanford Dependencies; “I like him”: nsubj(like, I), dobj(like, him), root(ROOT, like).
  13. What is WordNet? (s.59-60) Largest English lexical database: four databases (noun, verb, adjective, adverb) of synsets (sets of synonyms), e.g. {musician, instrumentalist, player}.
  14. Which tools are commonly used instead of building models from scratch? (s.57) NLTK, Stanford CoreNLP, Spacy, Textblob (about 80% of applications).

Common confusions

  • Stem vs lemma: stem = crude chop, may be a non-word (exampl); lemma = dictionary headword (be, reapply) (s.31, s.34).
  • Stemming vs “inflectional stemming”: the latter only normalises grammatical variants (plural, tense) and is called morphological analysis (s.33).
  • ”.” is not always end of sentence: Ph.D., www.google.com, i.e., 4.3, Inc. (s.38, s.41); ? and ! are treated as unambiguous (s.37).
  • POS tagging vs parsing: POS tagging labels each word with a category (NN, VBZ…); parsing builds a full structure (tree/dependencies) connecting all words (s.46, s.53).
  • Phrase structure tree vs dependency graph: s.54 is a constituency-style tree (S, NP, VP, PP); s.55-56 show word-to-word grammatical relations (nsubj, dobj, nmod) (s.54-56).
  • WordNet word forms vs senses: unique forms < number of senses in each category (e.g. noun 94474 forms, 116317 senses) because words are polysemous (s.60).
  • Lesk on slide vs Lesk in NLTK: slide s.50 hand-computes overlap for two senses; NLTK returns savings_bank.n.02 for “I went to the bank to deposit money” (s.51); note that this is the financial container sense, not the institution sense. UNCLEAR (slide 51): the slide does not comment on why that sense is chosen.

Added context

The “Simplified Lesk” formula written under Formulas is a notation of the slide’s verbal description, not a slide formula. The statement that Lancaster is “more aggressive” than Porter is read off the s.35 outputs; the slide gives no commentary.

Slide coverage

SlideContentCovered in
31Lemmatization examples, CoreNLPDetailed notes 1; Figures
32Morphology: morphemes, stems, affixesDetailed notes 2; Figures
33Inflectional stemming, ambiguity (bored)Detailed notes 3
34Stemming = crude chopping, automat; compress exampleDetailed notes 3; Figures
35Stemming with NLTK: Porter vs LancasterDetailed notes 3; Figures
36Divider: Sentence Boundary DetectionDetailed notes 4
37Sentence boundary determination: why, ? ! vs .Detailed notes 4
38Sentence segmentation, binary classifierDetailed notes 4
39EOS decision treeDetailed notes 4; Figures
40Implementing decision treesDetailed notes 4
41Hand-written EOS algorithm, examplesDetailed notes 4; Figures
42Divider: Syntactic AnalysisDetailed notes 5
43Syntactic analysis: POS, WSD, parsingDetailed notes 5
44POS tagging: motivation, categoriesDetailed notes 6
45POS uses: feature reduction, genreDetailed notes 6
46Penn Tree Bank, Brill, Stanford ParserDetailed notes 6
47Penn Tree Bank POS set tableDetailed notes 6; Figures
48Stanford Parser online demoDetailed notes 6; Figures
49WSD of “bank”, two WordNet sensesDetailed notes 7; Figures
50Simplified Lesk algorithmDetailed notes 7; Figures
51NLTK Lesk, synsets of bankDetailed notes 7; Figures
52BabelfyDetailed notes 7; Figures
53Parsing: definition, tree, whyDetailed notes 8; Figures
54Phrase structure tree, Johnson/XYZ/SmithDetailed notes 8; Figures
55Universal dependencies: “I like him”Detailed notes 8; Figures
56Universal dependencies: Johnson sentenceDetailed notes 8; Figures
57Off-the-shelf tools (80%)Detailed notes 9
58Divider: WordNetDetailed notes 10
59WordNet: 4 databases, synsetsDetailed notes 10; Figures
60WordNet size tableDetailed notes 10; Figures

Previous: 02 Pre-processing - Chunk 01 (s.1-30) | Next: 02 Pre-processing - Chunk 03 (s.61-90) | Up: 02 Pre-processing - Index

Graph View

Backlinks