Words and Meaning - LTHfileadmin.cs.lth.se/cs/Education/EDA171/Slides/EDA171... · Web viewWords...
-
Upload
vuongthien -
Category
Documents
-
view
220 -
download
6
Transcript of Words and Meaning - LTHfileadmin.cs.lth.se/cs/Education/EDA171/Slides/EDA171... · Web viewWords...
Language Processing and Computational Linguistics
EDA171/DATN06 – Lecture 13Lexical Semantics
Richard [email protected]
http://www.cs.lth.se/~richard
© Pierre Nugues, Lecture 13, September 30, 2008 1
Today’s Lecture
Lexical Semantics: how do we represent and process the meaning of words? The structure of dictionaries. Ontologies / semantic networks. Word senses. Sense ambiguity. Sense disambiguation. Putting pieces together: valence, selectional
restrictions, …
© Pierre Nugues, Lecture 13, September 30, 2008 2
Words and Meaning
Referred to as lexical semantics: Classes of words: If it is hot, can it be cold? Definition What is a meal? What is table? Reasoning: The meal is on the table. Is it cold?
© Pierre Nugues, Lecture 13, September 30, 2008 3
Categories of Words
Expressions, which are in no way composite, signify substance, quantity, quality, relation, place, time, position, state, action, or affection. To sketch my meaning roughly, examples of substance are ‘man’ or ‘the horse’, of quantity, such terms as ‘two cubits long’ or ‘three cubits long’, of quality, such attributes as ‘white’, ‘grammatical’. ‘Double’, ‘half’, ‘greater’, fall under the category of relation; ‘in the market place’, ‘in the Lyceum’, under that of place; ‘yesterday’, ‘last year’, under that of time. ‘Lying’, ‘sitting’, are terms indicating position, ‘shod’, ‘armed’, state; ‘to lance’, ‘to cauterize’, action; ‘to be lanced’, ‘to be cauterized’, affection.
Aristotle, Categories, IV. (trans. E. M. Edghill)© Pierre Nugues, Lecture 13, September 30, 2008 4
Representation of Categories
substance quality state relation action position place time quantity affection
expressions
© Pierre Nugues, Lecture 13, September 30, 2008 5
Definitions
Short texts describing a word: A genus or superclass using a hypernym. Specific attributes to differentiate it from other
members of the superclass. This part of the definition is called the differentia specifica.
bank (1.1): a land sloping up along each side of a canal or a river.hedgehog: a small animal with stiff spines covering its back.waiter: a person employed to serve customers at their table in a restaurant, etc .
© Pierre Nugues, Lecture 13, September 30, 2008 6
Relations between Words in a Dictionary
Synonymy/Antonymy Hyponyms/Hypernyms is_a(tree, plant), life form, entity Meronyms/Holonyms part_of(leg, table) …
© Pierre Nugues, Lecture 13, September 30, 2008 7
Semantic Networks – Ontologies
food
animates
human beings furnitureanimals
substance
possess
eat
carnivoresinsectivores
meatmammals
eat
© Pierre Nugues, Lecture 13, September 30, 2008 8
An Example: WordNet
Based on the concept of the synset.Nouns Hyponyms/hypernyms
Synonyms/antonymsMeronyms/holonyms
Adjectives Synonyms/antonymsRelational fraternal --> brother
Verbs Semantic domains (body function, change, communication, perception, contact, motion, creation, possession, competition, emotion, cognition, social interaction, weather)Synonymy, Antonymy: (rise/fall, ascent/descent, be born/die)“Entailment”: succeed/try, snore/sleep
© Pierre Nugues, Lecture 13, September 30, 2008 9
Lexical Database
%% is_a(?Word, ?Hypernym)is_a(hedgehog, insectivore).is_a(cat, feline).is_a(feline, carnivore).is_a(insectivore, mammal).is_a(carnivore, mammal).is_a(mammal, animal).is_a(animal, animate_being).
hypernym(X, Y) :- is_a(X, Y).hypernym(X, Y) :- is_a(X, Z), hypernym(Z, Y).
© Pierre Nugues, Lecture 13, September 30, 2008 10
Semantics and Reasoning
The caterpillar ate the hedgehog.Representation:∃(X, Y), caterpillar(X) & hedgehog(Y) & ate(X, Y).
Reasoning (inference):It is untrue because the query:?- predator(X, hedgehog)X = foxes, eagles, car drivers, …but no caterpillar.
© Pierre Nugues, Lecture 13, September 30, 2008 11
Word ambiguity: Homonymy and Polysemy
Words are ambiguous: A single form may have more than one entry (homonymy) and sense (polysemy).The Oxford Advanced Learner’s Dictionary (OALD) lists five entries for bank: Raised ground Organization Turn.and five senses for the first entry.The distinction between homonymy and polysemy is sometimes problematic. WordNet does not distinguish.Polysemy is productive! “bank of English”
© Pierre Nugues, Lecture 13, September 30, 2008 12
Sense Distinctions Vary Across Languages
Granularity: Boundaries:Swedish English French Welshsätta2 gwyrdd
ställa1 put1 vert
lägga1 bleuglas
grisllwyd
brun
© Pierre Nugues, Lecture 13, September 30, 2008 13
Sense Tagging Using the OALD
The patron ordered a mealWords Definitions Sensepatron Correct sense: A customer of a shop, restaurant, theater 1.2
Alternate sense: A person who gives money or support to a person, an organization, a cause or an activity
1.1
ordered Correct sense: To request somebody to bring food, drink, etc in a hotel, restaurant etc.
2.3
Alternate senses: To give an order to somebodyTo request somebody to supply or make goods, etc.To put something in order
2.12.22.4
Meal Correct sense: The food eaten on such occasionAlternate sense: An occasion where food is eaten
1.21.1
© Pierre Nugues, Lecture 13, September 30, 2008 14
Identifying Senses Automatically
The problem of determining sense is called word sense disambiguation.This is one of the most difficult tasks in NLP – the best systems are only slightly better than the baseline.Two broad approaches: Statistical methods based on co-occurrence statistics Dictionary-based methods
Hybrids are common.Both approaches rely crucially on the context to disambiguate.
© Pierre Nugues, Lecture 13, September 30, 2008 15
Bayesian Sense Disambiguation
We analyze the interaction between bank and market finance in a model where bankers gather information through monitoring…Statistical techniques optimize a sequence of semantic tags.The context
€
C is defined as
€
w−m ,w−m +1,...,w−1,w,w1,...,wm−1,wm .The optimal sense
€
ˆ s = argmaxsi ,1≤ i≤n
P(si | C).Using Bayes’ theorem, we have:
)|,,...,,,...,,()(maxarg
)|()(maxargˆ
11111,
1,
immmminis
iinis
swwwwwwPsP
sCPsPs
i
i
−−+−−≤≤
≤≤
=
=
© Pierre Nugues, Lecture 13, September 30, 2008 16
Naïve Bayes
The Naïve Bayes model is based on the bag-of-word approach: it assumes that words occur independently of each other. We replace )|,,...,,,...,,( 1111 immmm swwwwwwP −−+−− with the product of probabilities:
∏≠−=≤≤
=m
jmjiji
nisswPsPs
i 0,1,)|()(maxargˆ
To train the model, we can use SemCor, which is a sense-annotated corpus for English.In addition to supervised methods like Naïve Bayes, semisupervised and unsupervised algorithms exist.
© Pierre Nugues, Lecture 13, September 30, 2008 17
Using Dictionaries (Lesk and derived methods)
We analyze the interaction between bank and market finance in a model where bankers gather information through monitoring and screeningMaximally overlapping definitions:Bank: Bank1: The land sloping up along each side of a river or a canal; he
ground near a river Bank2: An organization or a place that provides a financial service.
Customers keep their money in the bank safely and it is paid out…Finance: Finance: The money used or needed to support an activity, project,
etc; the management of money
© Pierre Nugues, Lecture 13, September 30, 2008 18
Valence Patterns
Dictionaries store information about how words combine with other words to form larger structures. This information is called valence (cf. valence in chemistry)
OALD, tell1:~ sb (sth) / ~ sth to sb
I told him a lie I told a lie to him
© Pierre Nugues, Lecture 13, September 30, 2008 19
Syntactic Side: Verb Construction Models
English French Germandepend + on + object noun group
dépendre + de + object noun group
hängen + von + dative noun group + ab
I like + verb-ing (gerund)
Ça me plaît de + infinitive
es gefällt mir + zu + infinitive
require + verb-ing (gerund)
demander + de + infinitive
verlangen + accusative noun group
© Pierre Nugues, Lecture 13, September 30, 2008 20
Semantic Side: Selectional Restrictions
Three kinds of wanting:1. Wanting something to happen,2. Wanting an object,3. Wanting a person.and (2.) will be mapped on:word(category: verb, agent: persons, object: objects) --> [want].Properties of word mean: adjective, qualify only persons, and express badness:word(category: adjective, applyTo: persons, expresses: badness)--> [mean].
© Pierre Nugues, Lecture 13, September 30, 2008 21
Case Grammar
An idea from Indian grammarians (~500 BC): Define a set of semantic cases (or semantic roles):•An Agent – Instigator of the action (typically animate)•An Instrument – Cause of the event or object in causing the event
(typically animate)•An Object – Entity that is acted upon or that changes, the most
general case.•A Dative – Entity affected by the action (typically animate)•A Locative – Place of the event•…Used by linguists to study the relations between valence patternsUsed in NLP to build flexible template-filling applications
© Pierre Nugues, Lecture 13, September 30, 2008 22
Case Grammar: An Example
open(O, {A}, {I})
The door opened O = doorJohn opened the door O = door and A = JohnThe wind opened the door O = door and I = windJohn opened the door with a chisel
O = door, A = John, and I = chisel
© Pierre Nugues, Lecture 13, September 30, 2008 23
Parsing with Cases
The waiter brought the meal to the patronIdentify the verb bring and apply constraints:
Bring
Case Value
Agentive Animate (Obligatory) The waiter
Object (Obligatory) the mealDative Animate (Optional) the
patronTime (Obligatory) past
© Pierre Nugues, Lecture 13, September 30, 2008 24
FrameNet
In 1968, Fillmore wrote an oft-cited paper on case grammars.Later, he started the FrameNet project.In FrameNet, Fillmore no longer uses universal cases but a set of frames – semantic templates – where each frame is specific to a class of words.Framenet is a medium-sized lexical database that maps English words to frames.Examples of frames: Revenge, Statement, Impact, …
© Pierre Nugues, Lecture 13, September 30, 2008 25
The Revenge Frame
15 lexical units (dictionary entries):avenge.v, avenger.n, get back (at).v, get_even.v, retaliate.v, retaliation.n, retribution.n, retributive.a, retributory.a, revenge.n, revenge.v, revengeful.a, revenger.n, vengeance.n, vengeful.a, and vindictive.a.Examples of frame elements (template slots)Avenger, Punishment, Offender, Injury, and Injured_party.The word in a sentence that is tied to a frame is called the target.
© Pierre Nugues, Lecture 13, September 30, 2008 26
Annotation
1. [<Avenger> His brothers] avenged [<Injured_party> him].2. [<Punishment>With this], [<Avenger> El Cid] at once avenged
[<Injury> the death of his son].3. [<Avenger> Hook] tries to avenge [<Injured_party> himself]
[<Offender> on Peter Pan] [<Punishment> by becoming a second and better father].
© Pierre Nugues, Lecture 13, September 30, 2008 27
Automatic Frame-semantic Analysis
Given a syntax tree and a target word Find the arguments, Determine the template slot of each argument
to l dhim a lie I
SBJ IO B J
OBJRO O T
© Pierre Nugues, Lecture 13, September 30, 2008 28
Classification of Semantic Arguments
The task of classifying semantic arguments can be modeled as a statistical classification problem.
What features are useful for this task? Examples: Grammatical function: subject, object, … Voice: “I told a lie” / “I was told a lie” Semantic classes: “I told him” / “the note told him” Semantic class usually not available: use word instead
© Pierre Nugues, Lecture 13, September 30, 2008 29
Example
Given a dependency tree:
to l dhim a lie I
SBJ IO B J
OBJRO O T
We select the three dependents of told:I: GF = Subject voice = Active word = Ihim: GF = Indirect object voice = Active word = himlie: GF = Direct object voice = Active word = lie
© Pierre Nugues, Lecture 13, September 30, 2008 30
When Semantics Is Too Difficult: “Semantic Grammar”
sentence --> npInsectivores, ingest, npCrawlingInsects.npInsectivores --> det, insectivores.npCrawlingInsects --> det, crawlingInsects.insectivores --> [mole].insectivores --> [hedgehog].ingest --> [devours].ingest --> [eats].crawlingInsects --> [worms].crawlingInsects --> [caterpillars].det --> [the].
© Pierre Nugues, Lecture 13, September 30, 2008 31
An Application: EVAR
EVAR was a German project to provide information on trains.
noun
concrete abstract
thing location animate worth classifying time
transport human beast
© Pierre Nugues, Lecture 13, September 30, 2008 32
EVAR’s Valence Dictionary
1. fahren1.1(”The train is going from Hamburg to Munich”) Instrument: noun group (nominative), Transport, obligatory Source: prepositional group (Origin), Location, optional Goal: prepositional group (Direction), Location, optional
2. fahren1.2(”I am going by train from Hamburg to Munich”) Agent: noun group (nominative), Animate, obligatory Instrument: prepositional group (prep = mit), Transport,
optional Source: prepositional group (Origin), Location, optional Goal: prepositional group (Direction), Location, optional
3. Abfahrt1.1(”the departure of the train at Hamburg for Munich”) Object: noun group (genitive), Transport, optional
© Pierre Nugues, Lecture 13, September 30, 2008 33
Another Application: Traffic Accidents
[<Location>På landsvägen mellan Lund och Staffanstorp] krockade [<Time>i morse] [<Actor>en motorcyklist] [<Manner> våldsamt] [<Victim>med ett träd].
… efter en [<Manner>oförsiktig] omkörning …
© Pierre Nugues, Lecture 13, September 30, 2008 34