Thesauruses for Natural Language Processing Adam Kilgarriff Lexicography MasterClass and University of Brighton
Outline Definition Uses for NLP WASPS thesaurus web thesauruses Argument: words not word senses Evaluation proposals Cyborgs
a resource that groups words according to similarity What is a thesaurus? a resource that groups words according to similarity
Manual and automatic Manual Automatic Are they comparable? Roget, WordNets, many publishers Automatic Sparck Jones (1960s), Grefenstette (1994), Lin (1998), Lee (1999) aka distributional two words are similar if they occur in same contexts Are they comparable?
Thesauruses in NLP sparse data
Thesauruses in NLP sparse data does x go with y? New question: don’t know, they have never been seen together New question: does x+friends go with y+friends indirect evidence for x and y thesaurus tells us who friends are “backing off”
Relevant in: Parsing Bridging anaphors Text cohesion PP-attachment conjunction scope Bridging anaphors Text cohesion Word sense disambiguation (WSD) Speech understanding Spelling correction
Speech understanding He’s as headstrong as an alleg***** in the upwaters of the Yangtze
Speech understanding He’s as headstrong as an alleg***** in the upwaters of the Yangtze allegory?
Speech understanding He’s as headstrong as an alleg***** in the upwaters of the Yangtze allegory? alligator?
Speech understanding He’s as headstrong as an alleg***** in the upwaters of the Yangtze allegory? in upwaters? No alligator? in upwaters? No
Speech understanding He’s as headstrong as an alleg***** in the upwaters of the Yangtze allegory? in upwaters? No alligator? in upwaters? No allegory+friends in upwaters? No alligator+friends in upwaters? Yes
PP-attachment investigate stromatolite with microscope/speckles microscope: verb attachment speckles: noun attachment inspect jasper with spectrometer which?
PP attachment (cont) compare frequencies of <inspect, with, spectrometer> <jasper, with, spectrometer>
PP attachment (cont) compare frequencies of both zero? Try <inspect, with, spectrometer> <jasper, with, spectrometer> both zero? Try <inspect+friends, with, spectrometer+friends> <jasper+friends, with, spectrometer+friends>
Conjunction scope Compare old boots and shoes old boots and apples
Conjunction scope Compare Are the shoes old? old boots and shoes old boots and apples Are the shoes old?
Conjunction scope Compare Are the shoes old? Are the apples old? old boots and shoes old boots and apples Are the shoes old? Are the apples old?
Conjunction scope Compare Are the shoes old? Are the apples old? old boots and shoes old boots and apples Are the shoes old? Are the apples old? Hypothesis: wide scope only when words are similar
Conjunction scope Compare Are the shoes old? Are the apples old? old boots and shoes old boots and apples Are the shoes old? Are the apples old? Hypothesis: wide scope only when words are similar hard problem: thesaurus might help
Bridging anaphor resolution Maria bought a large apple. The fruit was red and crisp. fruit and apple co-refer
Bridging anaphor resolution Maria bought a large apple. The fruit was red and crisp. fruit and apple co-refer How to find co-referring terms?
Text cohesion words on same theme change in theme of words same segment change in theme of words new segment same theme: same thesaurus class
Word Sense Disambiguation (WSD) pike: fish or weapon We caught a pike this afternoon probably no direct evidence for catch pike probably is direct evidence for catch {pike,carp,bream,cod,haddock,…}
WordNet, Roget widely used for all the above
The WASPS thesaurus POS-tag, lemmatise and parse the BNC (100M words) credit: David Tugwell EPSRC grant K8931 POS-tag, lemmatise and parse the BNC (100M words) Find all grammatical relations <obj, climb, bank> <modifier, big, bank> <subject, bank, refuse> 70 million triples
WASPS thesaurus (cont) Similarity: <obj, drink, beer> <obj, drink, wine> one point similarity between beer and wine count all points of similarity between all pairs of words weight according to frequencies product of MI: Lin (1998)
Word Sketches one-page summary of a word’s grammatical and collocational behaviour demo: http://wasps.itri.bton.ac.uk the Sketch Engine input any corpus generate word sketches and thesaurus just available now
Nearest neighbours to zebra
Nearest neighbours zebra: giraffe buffalo hippopotamus rhinoceros gazelle antelope cheetah hippo leopard kangaroo crocodile deer rhino herbivore tortoise primate hyena camel scorpion macaque elephant mammoth alligator carnivore squirrel tiger newt chimpanzee monkey
exception: exemption limitation exclusion instance modification restriction recognition extension contrast addition refusal example clause indication definition error restraint reference objection consideration concession distinction variation occurrence anomaly offence jurisdiction implication analogy pot: bowl pan jar container dish jug mug tin tub tray bag saucepan bottle basket bucket vase plate kettle teapot glass spoon soup box can cake tea packet pipe cup
VERBS measure determine assess calculate decrease monitor increase evaluate reduce detect estimate indicate analyse exceed vary test observe define record reflect affect obtain generate predict enhance alter examine quantify relate adjust boil simmer heat cook fry bubble cool stir warm steam sizzle bake flavour spill soak roast taste pour dry wash chop melt freeze scald consume burn mix ferment scorch soften
ADJECTIVES hypnotic haunting piercing expressionless dreamy monotonous seductive meditative emotive comforting expressive mournful healing indistinct unforgettable unreadable harmonic prophetic steely sensuous soothing malevolent irresistible restful insidious expectant demonic incessant inhuman spooky pink purple yellow red blue white pale brown green grey coloured bright scarlet orange cream black crimson thick soft dark striped thin golden faded matching embroidered silver warm mauve damp
Nearest neighbours crane winch swan heron tern mast gull tractor rigging truck pump curlew flamingo
no clustering (tho’ could be done) no hierarchy (tho’ could be done) rhythm all on the web: http://wasps.itri.bton.ac.uk registration required
The web an enormous linguist’s playground Computational Linguistics Special Issue, Kilgarriff and Grefenstette (eds) 29 (3) (coming soon)
Google sets http://labs.google.com/sets Input: zebra giraffe buffalo
Google sets http://labs.google.com/sets Input: zebra giraffe buffalo kudu hyena impala leopard hippo waterbuck elephant cheetah eland
Google sets http://labs.google.com/sets Input: harbin beijing nanking
Google sets http://labs.google.com/sets Input: harbin beijing nanking Output: shanghai chengdu guangzhou hangzhou changchun zhejiang kunming dalian jinan fuzhou
Tree structure Roget all human knowledge as tree structure 1000 top categories subdivisions like this etc
Directories and thesauruses Yahoo, http://www.yahoo.com Open directory project, http://dmoz.org all human activity as tree structure plus corpus at every node gather corpus, identify domain vocabulary Gonzalo and colleagues, Madrid, CL Special Issue Agirre and colleagues, ‘topic signatures’
Words and word senses automatic thesauruses words
Words and word senses automatic thesauruses manual thesauruses words simple hierarchy is appealing homonyms
Words and word senses automatic thesauruses manual thesauruses words simple hierarchy is appealing homonyms “aha! objects must be word senses”
Problems Theoretical Practical
Theoretical
Wittgenstein Don’t ask for the meaning, ask for the use
Practical
Problems Practical . a thesaurus is a tool if the tool organises words senses you must do WSD before you can use it WSD: state of the art, optimal conditions: 80% .
Problems Practical a thesaurus is a tool if the tool organises words senses you must do WSD before you can use it WSD: state of the art, optimal conditions: 80% “To use this tool, first replace one fifth of your input with junk”
Avoid word senses
Avoid word senses This word has three meanings/senses
Avoid word senses This word has three meanings/senses This word has three kinds of use well founded empirical we can study it
sorry, roget
sorry, AI
sorry, AI AI model for NLP: NLP turns text into meanings AI reasons over meanings word meanings are concepts in an ontology a Roget-like thesaurus is (to a good approximation) an ontology Guarino: “cleansing” WordNet If a thesaurus groups words in their various uses (not meanings) not the sort of thing AI can reason over
sorry, AI “linguistics expressions prompt for meanings rather than express meanings” Fauconnier and Turner 2003 It would be nice if … But …
Evaluation manual thesauruses automatic thesauruses: attempts not done pseudo-disambiguation (Lee 1999) with ref to manual ones (Lin 1998)
Task-based evaluation
Task-based evaluation Parsing PP-attachment conjunction scope Bridging anaphors Text cohesion Word sense disambiguation (WSD) Speech understanding Spelling correction
What is performance at the task with no thesaurus with Roget with WordNet with WASPS
Plans set up evaluation tasks theseval web-based thesaurus campaign Open Directory Project hierarchies campaign
Cyborgs Robots: will they take over? Rod Brooks’s answer: Wrong question: greatest advances are in what the human+computer ensemble can do
Cyborgs A creature that is partly human and partly machine Macmillan English Dictionary
Cyborgs and the Information Society The thedsaurus-making agent is part human (for precision), part computer (for recall).
Summary: Thesauruses for NLP Definition Uses for NLP WASPS thesaurus web thesauruses Argument: words not word senses Evaluation proposals Cyborgs
Thesaurus-makers of the future?