Concluding

1
Concluding

• NLG as a part of Language Technology

2
NLG

• Convert computational input into nice output
  (digestible by human user)
• nice output can include other modalities as
  well as text
• whats nice for one user is not necessarily nice
  for another. (Think of differences in reading
  skills, expertise, etc. also, disabilities)
• computational databases, logic, ...
• Were talking HCI

3
NLU NL Understanding

• Convert text into computational output
• This is the inverse of NLG
• As a result, NLG and NLU are similar yet very
  different

4
Differences NLG/NLU

• NLG involves choosing between ways in which
  information can be expressed. E.g.,
• Id like to buy the green chair
• Id like to buy the chair that costs 50
• Id like to buy the cheap chair
• Id like to buy the chair in the corner
• Id like to buy that chair

5
Differences NLG/NLU

• NLU has to cope with this same range of
  expressions ... as possible inputs
• Allowing such open-ended input poses a huge
  problem for NLU
• Many present systems finesse this problem by
  limiting the users input possibilities
• Example WYSIWYM (previous lecture)

6
Connections NLG/NLU

• 1. Many systems combine the two. E.g.,
• Most dialogue systems require understanding of NL
  as well as generation of NL. (Exception NECA)
• Future MT and summarisation systems too.

7
Connections NLG/NLU

• 2. An NLG system should not generate text that an
  NL system could not understand.
• Input 1 man(x) old(x) gt entitled(x)
• woman(x) old(x) gt entitled(x)
• Output Old men and women are entitled to benefits

8
Connections NLG/NLU

• Input 2 man(x) old(x) gt entitled(x)
• woman(x) old(x) gt entitled(x)
• Output2 Old men and women are entitled to
  benefits
• Two inputs having the same output
• Ambiguity (assignment 1)

9
In sum

• Weve taught you about Language Technology
• Weve focussed on NLG
• This should equip you with the knowledge to find
  out about NLU systems as well.