computational_semantics.md

Word Relations

Relations

lemma and wordform

• lemma
  • same stem, pos, rouch semantics
  • eg, bank, sing
  • a lemma can have senses
• wordform
  • the inflected word as seen in the text
  • eg, banks, sung

Homonymy

share a form but have unrelated, distinct meanings

Polysemy

polysemous words have related meanings

eg, bank can mean a financial institution or a building belonging to the financial institution

Metonymy or Systematic Polysemy

eg, school, university, hospital all can mean the institution or the building

Synonymy

• relation between senses and not words

Antonymy

• relation between senses

Hypo and Hypernymy

• ...

Meronymy

• ...

WordNet

• hierarchically organised lexical database
• online thesaurus + aspects of a dictionary

Definition of sense

• using synset (synonym set), set of near-synonyms, gloss

Word Sense Disambiguation

• Given
  • a word in context
  • a fixed inventory of potential word senses
• Decide
  • which sense of the word this is

Two variations

• Lexical Sample Task

  • small pre-selected set of target words
  • supervised machine learning

• All-words Task

  • every word in an entire text

Methods

Supervised Machine Learning

• training corpus is used to train a classifer
• what do we need?
  • a tag set
  • the training corpus
  • set of features extracted from the corpus
  • classifier

Feature Vectors

• vectors of sets of feature/value pairs
• represented as an ordered list of values
• these vecs represent for eg, the window of words around the target
• 2 types of features
  • collocational
    • features about words at specific positions around the target
  • bag-of-words
    • features about words that occur anywhere in the window

Example

An electric guitar and bass player stand off to one side and not really part of the scene

Assume an window of +/-2 from the target. The window contains ['guitar', 'and', 'player', 'stand']

Collocational features will have positional information.

Bag-of-words contains and vocabulary and a binary indicator. Assume the vocab to be ['fishing', 'big', 'sound', 'player', 'fly', 'rod', 'pound', 'double', 'runs', 'playing', 'guitar', 'band'] The vector for guitar and bass player stand: [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0]

Classification

• Input:
  • a word w and some features f
  • a fixed set of classes c = {c1, c2, ...}
• Output:
  • a predicate class c belongs to C

Types

• Naive Bayes
  • classification based on Bayes Rule
  • relies on simple representations of documentation like bag-of-words
• Logistic regression
• Neural Nets
• ...

Dictionary Methods

Simple Lesk Algorithm

the bank can guarantee deposits will eventually cover future tuition costs because it invests in adjustable-rate mortgage securities

• given the two WordNet senses
  • financial institution
  • sloping land
• words like deposits and mortgage can be found in the WordNet related to the first sense

Corpus Lesk Algorithm

• needs some sense-labelled data (like SemCor)
• take all the sentences with the relevant word sense
• add these to the gloss+examples for each sense, call it signature of the sense
• choose sense with most overlap between context and signature
• weigh each overlapping word by inverse document frequency

Graph-based Methods

• WordNet as graph
  • senses are nodes
  • relations (hypernymy, etc) are edges
  • add edges between word and unambiguous gloss words
• Insert target word or words in its sentential context into the graph, with directed edges to their senses
• Choose the most central sense

Semi-supervised Learning

• Supervised and dictionary-based approaches require large hand-built resources
• What if you don’t have so much training data?
• Solution:
  • Generalize from a very small hand-labeled seed-set

Semantic Role Labelling