Exhibit 07Classification

Part-of-Speech Tagging

Every word in a sentence gets one of 49 grammatical tags. It is the textbook classification problem with real, human-annotated answers, and the kind of many-label choice Jev is built for.

In scopeMany narrow classification judgments, each choosing from dozens of labels: Jev's headline use case. TypeSafe launch blog: “Jev supports a cardinality up to 255.
World
State
Decision
Action
New state
1 / 200

There are currently 5 locations:

gold = human annotationjevhmm = textbook HMM + Viterbimft = most frequent tag
This session

Tag a few test sentences to build up accuracy against the human annotation.

On all 2,593 tokens of the 200 test sentences here, the textbook taggers score 92% (HMM) and 88% (most frequent tag), trained on the UD English EWT r2.18 (commit b7711cc) training split.

Tag families
nounsverbsadjectivesadverbspronounsdeterminersprepositions & conjunctionsnumberspunctuationother
All 49 tags
  • NNNoun, singular or mass: dog, water, idea
  • NNSNoun, plural: dogs, ideas
  • NNPProper noun, singular: Google, London, Monday
  • NNPSProper noun, plural: Americans, Alps
  • VBVerb, base form: to go, can see, please help
  • VBDVerb, past tense: went, saw, morphed
  • VBGVerb, gerund or present participle: going, seeing
  • VBNVerb, past participle: gone, seen, has taken
  • VBPVerb, present tense, not 3rd person singular: I go, they are
  • VBZVerb, present tense, 3rd person singular: she goes, it is
  • MDModal verb: can, will, should, might
  • JJAdjective: big, green, important
  • JJRAdjective, comparative: bigger, better
  • JJSAdjective, superlative: biggest, best
  • RBAdverb: quickly, not, very, also
  • RBRAdverb, comparative: faster, more (as adverb)
  • RBSAdverb, superlative: fastest, most (as adverb)
  • RPParticle of a phrasal verb: give UP, turn OFF
  • WRBWh-adverb: where, when, why, how
  • PRPPersonal pronoun: I, you, he, it, them
  • PRP$Possessive pronoun before a noun: my, your, their
  • WPWh-pronoun: who, what, whom
  • WP$Possessive wh-pronoun: whose
  • EXExistential there: THERE is a problem
  • DTDeterminer: the, a, this, some, no
  • PDTPredeterminer before a determiner: ALL the, BOTH his
  • WDTWh-determiner: which, that (introducing a relative clause)
  • INPreposition or subordinating conjunction: in, of, if, because
  • TOThe word to, as infinitive marker or preposition
  • CCCoordinating conjunction: and, but, or
  • POSPossessive ending: the 's in John's
  • CDCardinal number: 3, three, 2004, 1.5
  • LSList item marker: 1), a., (b) at the start of an item
  • .Sentence-final punctuation: . ! ?
  • ,Comma: ,
  • :Colon, semicolon, dash or ellipsis inside a sentence: : ; -- ...
  • ``Opening quotation mark: “ or " at the start of a quote
  • ''Closing quotation mark: ” or " at the end of a quote
  • -LRB-Opening bracket: ( [ {
  • -RRB-Closing bracket: ) ] }
  • HYPHHyphen joining parts of a split word: the - in e - mail
  • NFPSuperfluous or decorative punctuation: *** ~~ :)
  • $Currency symbol: $ £ €
  • SYMSymbol used as a word: % & = + /
  • UHInterjection: yes, oh, thanks, hello
  • FWForeign word: et, al, bon, voyage
  • ADDEmail address or URL: www.example.com
  • GWFragment of a word split by a typo: the 'every' of 'every thing'
  • AFXSeparated affix or prefix: pre, anti (as in pre - war)

Jev decision

idle
Press Tag with Jev. Each token becomes one question with all 49 tags as options; up to 16 questions go in one call. Click a token to see Jev's full distribution.
NN · Noun, singular or mass
NNS · Noun, plural
NNP · Proper noun, singular
NNPS · Proper noun, plural
Other (45)
Lesson · Classification
Word in context → one of 45 tags
Part-of-speech tagging is a long line of narrow classification decisions, each choosing from dozens of labels, which is what TypeSafe says Jev is for. Unlike the Restaurant exhibit, there is a right answer: every test sentence was tagged by human annotators (Universal Dependencies English EWT). The classic solutions are the textbook ones: tag each word with its most common tag in the training data, or run a hidden Markov model with the Viterbi algorithm so neighbouring tags inform each other. Both are trained on the same corpus; Jev is not trained on it and sees each sentence cold, one question per word.