The meaning of words

1
The meaning of words
2
Some aspects of a word

• Token
• Type
• Word form
• Base form
• Lemma
• Lexeme

3
Ambiguous words

• Homonymy
• Equal in form, different in meaning
• Homography
• Equal in orthographic form, different in meaning
• Homophony
• Equal in phonetic form, different in meaning
• Polysemy
• Equal in form, (partly) related meaning

4
Ambiguous words

• Homonyms with totally different syntactic
  distribution ? usually different POS
• Polysemic have the same POS
• Other homonyms

5
Meaning Relations to other lexemes

• Homonomy
• Polysemy
• Synonymy
• Hyponymy, hypernyms
• Other relations Antonyms, Part-of, Has-part,
  Member-of, Has-member,

6
Synonyms different lexemes with the same meaning

• Substitutable in all possible environments
• in some
• Difference in
• Polysemy
• Shades of meaning
• fare, price
• Collocation behaviour
• Register

7
Hyponymy relations

• mammal dog poodle
• Other relations
• car ?? wheel
• team ?? member
• good ?? bad
• Cygwin Wordnet

8
Meaning internal structure

• Semantic decomposition
• die kill murder
• darken
• kitten, puppy, calf, child
• mug, cup

9
Semantic roles

• The stone was thinking of breakfast
• She built a house
• The children were sleeping
• He gave him a gift
• AGENT, THEME, EXPERIENCER, RESULT, BENEFICIARY,
  INSTRUMENT,

10
Selectional restrictions

• Lexemes place restrictions on the lexemes that
  accompany them.
• A verb want subjects and/or objects to fill
  certain roles, and restrict the possible lexemes
  that can appear.

11
Meaning Co-occurrence

• You shall know a word by the company it keeps
• N-grams
• Phrases, sentences
• Syntactic relations, role frames
• Paragraphs
• Texts

12
Word sense disambiguation

• Machine-learning approaches
• Similarities to POS-tagging
• Target word context
• Feature vector N-gram
• More information larger window structured
  information

13
Random Indexing

• Unsupervised method for representing the meaning
  of a lexeme with its contexts.
• Magnus Sahlgren, SICS, Stockholm
• high-dimensional sparse random vectors to index
  words or documents
• Word meaning are represented by high-dimensional
  semantic context vectors

14
Random Indexing

• A unique random vector for every word
• A window
• Lemmatized words or tokens
• Add the vectors for the words in the window to
  the context vector
• Distance metrics between vectors
• Synonyms 60-70 correct at TOEFL test

15
Words and IR

• Bag-of-words
• Syntactic analysis could provide means to
  differentiate between car as AGENTS, INSTRUMENTS,
  THEMES, etc. Why not?
• Stemming
• Lemmatizing

16
Word sense problems and IR

• Query expansion
• Document clustering