14.13 Different types of word senses. Word Sense Disambiguition

Word Sense

• A word sense is a meaning for a word in the dictionary
• Monosemy: A word that has a single clear meaning in all contexts:
  • Table usually refers to a piece of furniture
• Polysemy: Multiple related meanings :
  • bank can refer to a financial institution or a side of a river,
  • foot is a part of body or the measuring
  • plain can mean clear unadorned or obvious
• Homonymy: A word with unrelated meanings:
  • bat can refer to the flying animal or a cricket bat
  • to lie (not tell truth or rest)
  • to bear (to give birth or to tolerate)

Word Sense Disambiguition

Word Sense Disambiguition (WSD) is the task of determining which of the possible senses of a ambiguous word is intended use of the word in the context

• This is done by looking at the context in which the word appears
• WSD assumes the word has finite number of discrete senses which are often provided by a dictionary
• The goal is to make a program to make a forced decision between the senses

Inputs and Evaluations for WSD:

• A dictionary which specifies the senses to be disambiguated
• A high annotated Test Corpus that has teh target or correct senses
  • Lexical Sample: used when a system needs to disambiguate a small sample of words
  • All-Words: '' '' '' all words in a piece of running textA

Challenges in WSD

• Polysemy: words having multiple related meanings hard to make the decision
• Context Sensitivity: A word's meaning can change based on context, syntactic structure and discourse context
• Ambiguity: senses may be closely related to eachouter or contextually dependent
• Data Sparsity: limited annotated data for WSD training especially for less frequent senses
• Domain Specificity: word meanings vary across domains require specialized knowledge

Approaches and Methods

• Dictionary based
• Supervised model
• semi supervised model
• unsupervised model

Dictionary based

• Lesk algorithm aims to choose the sense whose dict definition or gloss shares the most words with the target's neighborhood
• it gets all possible senses for the target word and then surrounding words from the dict
• then it measures the overlap between defs of the target word senses and the defs of the surrounding context words
• sense with the highest overlap score is selected
• Ex:
  • sense 1: "a financial establishment that accepts the deposits and offers loans"
  • sense 2 "A slope of land adjoining a river"
  • context words: "deposited", "money"
  • Overlap is greater for sense 1 for the adjoining words
  • Limitations: Dict entries can be short leading to insufficient overlaps, a solution like Corpus Lesk applies weights (IDF) to overlapping words

Supervised model

• methods that extract features from the text and train a classifier that assign the correct sense
• A feature is a numeric vector that encodes linguistic info as input to ML algorithms
• Collocation features: encode info about specific words or phrases in position specific relationships to the target words
• Bag of Words: represents the context of a target word as an unordered set of words ignoring their exact position

Bootstrapping methods:

• It trains a classifier on a seed set and uses it to label the unlabeled corpus, selects the most confident predictions, adds them to the training set

figure it out later dude