Semantic Resources

Short DIY clip by Iris Hendrickx.

Example by Noam Chomsky that illustrates the difference between syntax and semantic:

"Colorless green ideas sleep furiously"

From a syntactic viewpoint this sentence is grammatically correct; but it's completely unclear what this sentence means (its semantics).

For text mining applications it's a important challenge to recognize sentences with the same meaning, in particular in information extraction.

Semantic analysis: how do parts of a sentence relate to each other not semantically? This involves several tasks as several levels:

  • Words: word sense disambiguation
  • Sentences: semantic parsing, semantic role labelling
  • Discourse: text structure representation, co-reference resolution (find references to the same entity in a text), dialogue acts.

A clear difference with syntax is that meaning representations are not limited to sentence boundaries (e.g. see "discourse"), whereas syntactic analysis is limited to the grammatical role of words in a sentence.

Word senses

Semantic annotation at word level has to deal with:

  • Homonymy i.e. homophones or homographs
    • Two different things being spoken or written the same way
    • Head (leader) vs head (body part)
  • Polysemy
    • different but related meanings
    • bank: blood bank vs. national bank (but both store something valuable)
  • Vagueness: nurse (male) vs nurse (female)

Dealing with this requires disambiguation, using contextual and lexical information.

So for this we need a lexical-semantic resource!

A typical dictionary is an example! But there are a range of reasons why these are not sufficient:

  • Word definitions are circular
  • Who decides the meaning of a word?
  • Or how many meanings it has? Different dictionaries list different subsets.
  • Some meanings of words are more similar to each other than others; this is not distinguished clearly in dictionaries.

A better alternative is Wordnet, where words are defined as:

  • Synsets (collections of synonyms)
  • Hierarchical w.r.t. a set of hypernyms (superset) and one of hyponyms (subset)

Sentence level: semantic roles

Semantically a sentence is often regarded as communicating something about an event.

Example: "The girl kissed the boy"

We can distinguish thematic roles:

  • Agent (causer of event)
  • Theme (beneficiary of event)
  • Goal

Unfortunately, there is no agreement about a standard set of thematic roles. If you have a different theory about roles, you'll also end up with different annotations.

Also, because verbs restrict nouns in different ways, you cannot always apply the same rules for each sentence.

There are two different flavours of theory here:

  • Generalized semantic roles: you try to use the same set for all possible sentences; e.g. use the "agent" role in as many situations as possible
  • Specialized semantic roles: define specialized roles for verbs and events separately.

Generalized roles

E.g. generalized roles could be:

  • PROTO-AGENT
  • PROTO-THEME

where these roles have a set of characteristics. You can classify something as an agent when it shares enough of these characteristics. So the general case catches more specific cases by defining a broader set of characteristics and not requiring all to be present.

One example resource here is Propbank. This uses Predicate-Argument structures for semantic roles. Where these do not suffice, they also have some extra verb-specific roles.

Other resources:

  • FrameNet (Fillmore)
  • VerbNet (Kipper-Schuler, 2006)