Some questions (doubts) about STS (mostly from the perspective of distributional semantics)

1
Some questions (doubts) about STS(mostly from
the perspective of distributional semantics)

• Alessandro Lenci
• University of Pisa

2
What is semantic similarity?

• Is semantic similarity a well-grounded notion?
  Perhaps no...
• for concepts?
• cf. Goodmans argument that similarity is empty
  unless you specify similar with respect to what
• for lexical items?
• partially
• word similarity judgments hide many different
  types of semantic relations (cf. Rubinstein and
  Goodenough)
• for phrases and sentences?
• ??
• but we know some relations between sentences
  (e.g. entailment, contradiction, presupposition,
  etc.)

3
Which (basic) components for STS?

• Module for word-level semantic similarity
• Module for compositional semantic similarity
• do we know how to project semantic similarity
  from the word level to sentence or text level?

4
From semantic similarity to semantic relations
(in distributional semantics)

• Lexical entailment (cf. Dagan et al.)
• Classical semantic relations
• e.g. hypernymy, antonymy, etc.
• some words are very similar under one respect
  (e.g. increase / decrease, open / close, etc.)
  but express highly dissimilar concepts, etc.
• some texts can share most of their words, and
  still be very different
• Most students like to go dancing on Saturday.
• Few students like to go dancing on Saturday.
• Most students hate to go dancing on Saturday.

5
What to do?

• Prepare carefully designed and controlled data
  sets
• explore different dimensions of variation with
  respect to text similarity
• function words, lexical items, syntax, etc.
• explore different types of relations between
  texts covered by STS
• Analyse the factors that affect the judgments of
  STS