Semantics and Pragmatics of NLP Lexical Semantics: Polysemy · Semantics and Pragmatics of NLP...
Transcript of Semantics and Pragmatics of NLP Lexical Semantics: Polysemy · Semantics and Pragmatics of NLP...
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Semantics and Pragmatics of NLPLexical Semantics: Polysemy
Alex Lascarides
School of InformaticsUniversity of Edinburgh
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Outline
1 Why A Dictionary Won’t Do: Polysemy!
2 Different Kinds of Polysemy
3 How to Capture Lexical Generlisations
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Basic Lessons
The meanings of words sometimes predict aspects of theirsyntactic behaviour (regular polysemy).So lexical knowledge is the interface between worldknowledge and linguistic knowledge/processingDiscourse/ pragmatic processing interacts with lexicalsemanticsLexical semantic information can be modelled using atyped feature structure formalism, extended to handledefaults
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Introducing bake
(1) a. Kim is bakingb. The potatoes are bakingc. Kim is baking a caked. Kim is baking a cake for Sandye. Kim is baking Sandy a cakef. The clay bakedg. The clay baked hardh. Sandy was baking, sitting in the hot sun
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Sense enumeration: a strawman
(2) a. 〈bake, TakesNP, bake′1〉
b. 〈bake, TakesNP, bake′2〉
c. 〈bake, TakesNP.NP, bake′3〉
d. 〈bake, TakesNP.NP.PP, bake′4〉
e. 〈bake, TakesNP.NP.NP, bake′5〉
Senses connected by meaning postulates, such as:∀x , y [bake′
3(x , y) → bake′1(x) ∧ bake′
2(y)]
∀x , y [bake′4(x , y , z) → bake′
3(x , y)]
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Problems
Doesn’t capture generalisations (cf cook, paint)No distinction between homonyms (accidental polysemy)and related sensesMeaning postulates are unrestrictedCannot list all potential usages.
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Terminology
homonymy mogul :mound of snow vs. mogul :chinese emperorunpredictable polysemy bank :tilt of plane vs. bank :moundregular polysemy bank :building used by bank :institution
nonce not recorded in the lexicon, but interpretable viagenerative devices;The ham sandwich is getting impatient.
institutionalised recorded in the lexicon as derived by a regularprocess; teacher
lexicalised idiosyncratically augment or override regularlyderived information; in hospital.
established covers institutionalised and lexicalised
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
More Terminology
constructional polysemy a single sense assigned to a lexicalentry is contextually specialised
(3) a. That book is 500 pages longb. That book introduces syntax.c. That book is 500 pages long and
introduces syntax.
sense extension separate lexical entries are generated
(4) a. That chicken is healthy.b. That chicken is tasty.c.??That chicken is healthy and tasty.
Distinction between constructional polysemy and senseextension is not always clear cut.
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Back to the lexical entry for transitive bake
transitive-verbORTH : bake
SYN :
CAT : v
SUBJ :
phraseSYN:CAT : nSEM:INDEX : x
SUBCAT :
⟨ phraseSYN:CAT : nSEM:INDEX : y
⟩
SEM :
INDEX : e
LISZT :
⟨_bake3_relEVENT : eARG1 : xARG2 : y
⟩
HPSGish syntax and MRS — neutral wrt various approaches tolexical semantics
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
bread (NP)
phraseORTH : bread
SYN :
[CAT : nSUBCAT : 〈〉
]
SEM :
INDEX : w
LISZT :
⟨[_bread_relINST : w
]⟩
LISZTs are appended when signs are combined LISZT :
⟨_bake3_relEVENT : eARG1 : xARG2 : y
,
[_bread_relINST : y
]⟩ BAKE3(e, x , y) ∧ BREAD(y)
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Abbreviated description
transitive-verbORTH : bake
SYN :
CAT : vSUBJ : NP xSUBCAT : 〈NP y 〉
SEM :
INDEX : e
LISZT :
⟨_bake3_relEVENT : eARG1 : xARG2 : y
⟩
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Type for simple transitive verbs
transitive-verbORTH : string
SYN :
CAT : vSUBJ : NP xSUBCAT : 〈NP y 〉
SEM :
INDEX : e
LISZT :
⟨relEVENT : eARG1 : xARG2 : y
⟩
Generalisation: agents realised as subjects and patients asobjects.
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Type hierarchies>
���HHHsign
���HHH
word���
HHHverb noun
transitive-verb
bake3
phrase
relaaaaaaanoun_rel
���HHHphys_rel
���art_rel���
HHHphys_art_rel
verb_rel
tverb_rel
_bake3_rel
Information on higher types is inherited by lower types (andlexical entries, such as bake3)Multiple inheritance is possible
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Constructional polysemy: Logical Metonymy
(5) a. Sandy enjoyed the film/beer/hamburger.b. Sandy enjoyed the book.c. Sandy enjoyed reading the bookd. The goat really enjoyed your book.
Don’t want to enumerate a million senses of enjoy!It’s not a purely pragmatic phenomenon:??John enjoyed the doorstop. ??John enjoyed the tunnel.Generalisation: When NP is an artifact,enjoy NP means enjoy V-ing NP, where V is its purpose.But this can be overridden in sufficiently rich discoursecontexts.
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Capturing the Generalisation (Simplified):enjoy the book
book: enjoy: inherited info; begin, finish etc.2664bookSEM : book(y)
QUALIA :
"CONST : pagesTELIC : readAGENTIVE : write
#3775
266664coercing
CAT SUBCAT :
*24 npSEM : n [Q(y)]
QUALIA TELIC : act-on-pred P
35+
SEM : [e][enjoy(e, x, e′) ∧ act-on-pred/ P (e′, x, y) ∧ n ]
377775
enjoy the book:coercing
CAT SUBCAT :
⟨ npSEM : n book(y)
QUALIA TELIC : P read
⟩SEM : [e][enjoy(e, x , e′) ∧ / P read(e′, x , y) ∧ n book(y)]
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Constructional Polysemy: Sense Broadening
(6) a. There are clouds in the sky — it’s going to rain.b. There was a cloud of flies round the cow.c. The flies were pestering the horse — it swished its
tail at the buzzing cloud.
lex-count-nounORTH : cloudCAT : noun-catSEM : obj-noun-formula
QUALIA :
phys_obj /natural_obj
FORM :
nomformRELATIVE : indivABS. : amorph
CONSTITUENCY : phys_cum /water-vapour
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Sense extension
Grammatical effects: count -> mass, verbingEstablished and non-established sensesConventional nature of process
(7) a. Sandy drank a bottle of whisky.container -> contents
b. Sandy drank a bottleful of whisky.
(8) a. Kim ate some chicken.animal -> meat
b. That restaurant serves ostrich.
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
A particular kind of Sense Extension: ReferenceTransfer
(9) a. The ham sandwich has paid his check.physical object -> associated person
b. *The dark haired guy is in the microwave.
(10) Chester serves not just country folk, but farming,suburban and city folk too. You’ll see Armani driftinginto the Grosvenor Hotel’s exclusive (but exquisite)Arkle Restaurant and C+A giggling out of its streetfrontbrasserie next door. (Guardian Weekly)manufacturer -> product + clothes -> wearer
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
“Lexical Rules”
grinding < lexical-rulelex-count-nounORTH : 0SYN : noun-catSEM PRED : 3QUALIA : physical
→
lex-uncount-nounORTH : 0SYN : noun-catSEM PRED : grinding′( 3 )QUALIA : physical
meat-grinding < grinding[QUALIA : animal
]→
[QUALIA : edible_substance
]
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Verb classes: Levin (1993)
Large-scale descriptive account of (some) English verbs,pushing the idea that syntactic behaviour is (partly)semantically determined
alternationsThe ways in which arguments to verbs can be realiseddifferently.semantically coherent classes which exhibit the samealternationssome extended senses (eg whistle as a movement verb)
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
26 Verbs of Creation and Transformation
26.1 Build verbs Class members include: bake, cook
(11) Material/Product Alternation (transitive):a. Martha baked a loaf out of wholewheat flourb. Martha baked some wholewheat flour into a loaf
(12) Unspecified Object Alternation:a. Martha bakes breadb. Martha bakes
(13) Benefactive alternationa. Martha baked a loaf (out of wholewheat flour) for
her auntb. Martha baked her aunt a loaf (out of wholewheat
flour)
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
26.3 Verbs of Preparing
Class members include: bake, cook, fix, fry
(14) *Material/Product Alternation (transitive):a. ?Donna fixed a sandwich from last night’s leftoversb. *Donna fixed last night’s leftovers into a sandwich
(15) Benefactive alternationa. Donna fixed a sandwich for meb. Donna fixed me a sandwich
(16) *Causative alternationsa. Donna fixed a sandwichb. *a sandwich fixed
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
45 Verbs of change of state
45.3 Cooking verbs Class members include: bake, cook, fry
(17) Causative/Inchoative Alternationa. Jennifer baked the potatoesb. The potatoes baked
(18) Middle Alternationa. Jennifer baked Idaho potatoesb. Idaho potatoes bake beautifully
(19) Instrument Subject Alternationa. Jennifer baked the potatoes in the ovenb. This oven bakes potatoes well
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Representing alternations: Lexical rules2666666666666666664
creation-tverbORTH : 1
SYN :
2664CAT : vSUBJ : NP
xSUBCAT : 〈NP
y〉
3775
SEM :
2666664INDEX : e
LISZT :
*2
26664relEVENT : e
ARG1 : x
ARG2 : y
37775+
3777775
3777777777777777775
→
2666666666666666664
benef-PP-verbORTH : 1
SYN :
2664CAT : vSUBJ : NP
xSUBCAT : 〈NP
y, PP(for)
z〉
3775
SEM :
2666664INDEX : e
LISZT :
*2
26664relEVENT : e
ARG1 : x
ARG2 : y
37775,
24 benef_relEVENT : e
ARG1 : z
35+3777775
3777777777777777775
John baked a cake → John baked a cake for MaryAlex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Semi-productivity
??pig meaning meat, ??cow meaning meat,??John donated Oxfam a covenant etc.Avoid obscurity: use the form which has highest probability.
Estimating Probabilities: via Prob(lexical-entry | word-form)Seen lexical entries: Use frequencies in a very large corpus
(marked with senses!!)badger count noun, deer count noun.
Unseen lexical entries: Estimate the productivity of theappropriate lexical rule.badger meaning meat, deer meaning meat.
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Probabilities of Unseen Lexical Entries
1 Estimate the degree of productivity of a lexical rule lr bycomparing the number of attested outputs Mlr with thenumber of attested inputs Nlr seen in the corpus:
Prod(lr) =Mlr
Nlr
2 Use the Prod(lr)s to smooth over unseen data.
unseen-pr-mass(wf) =number-of-unattested-entries(wf)
freq(wf) + number-of-unattested-entries(wf)
est-freq(lei |wf) = unseen-pr-mass(wf)× Prod(lri)
ΣProd(lr1), . . . , Prod(lrn)
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Blocking
Blocking is an automatic consequence of avoid obscurity:
Prob(DEER-MEAT|venison) > Prob(DEER-MEAT|deer)
therefore generation of blocked forms is marked.
(20) a. That restaurant serves venison/?deer.b. There were five thousand extremely loud people
on the floor eager to tear into roast cow with bothhands and wash it down with bourbon whiskey.(Terry Pratchett)
Blocked forms dispreferred, but interpretable if otherpossibilities fail.Need formal account of pragmatic effects of unblocking
Alex Lascarides SPNLP: Lexical Polysemy
university-logo
DataDifferent Kinds of Polysemy
How to Capture Lexical Generlisations
Conclusions
Lexical semantics interacts in complex ways with syntaxand pragmatics.A dictionary model of the lexicon is too simplistic to do thisjustice.But manually constructing a lexical type hierarchy with richsemantic information is impractical.Can we use machine learning from corpora toautomatically acquire the lexical semantic information?
Next time: A case study—logical metonymy.
Alex Lascarides SPNLP: Lexical Polysemy