Natural Language Generation

1
Natural Language Generation

• Alice Oh
• aliceo_at_cs.cmu.edu
• 18 November 2014

2
What is NLG?

• Natural Language Understanding (NLU)
• Natural Language Generation (NLG)
• There has been active research in machine
  translation and automatic summarization (e.g.,
  stock quotes, medical information) communities.

Text
Semantic (Syntactic) Representation
Text
Semantic (Syntactic) Representation
3
Application of NLG in MT

• Generation of output
• Generation of paraphrase
• Generation of interlingua (SQL queries,
  summarization tables, etc.)
• Generation of controlled language
• The whole MT problem can be viewed as an NLG
  problem!

Source Language
Target Language
NLG
4
Why spend time/effort on NLG?

• What is the problem here? There are many more
  researchers working on NLU than NLG (esp. in the
  U.S.)
• Why? Because of the squeezing out toothpaste
  analogy -- which is not true!
• Why is NLG important? Its the only thing users
  see/hear (do not ruin your otherwise great
  system)
• What makes NLG difficult? NLG needs to KNOW vs.
  UNDERSTAND
• Existing comprehension systems as a rule extract
  considerably less information from a text than a
  generator must appreciate in generating one.
  Examples include the reasons why a given word or
  syntactic construction is used rather than an
  alternative, what constitutes the style and
  rhetoric appropriate to a given genre and
  situation, or why information is clustered in one
  pattern of sentences rather than another. -
  McDonald, 1993

5
Levels of NLG
6
Surface Realization

• Determining how the underlying content of a text
  should be mapped into a sequence of grammatically
  correct sentences. An NLG system has to decide
  which syntactic form to use, and it has to ensure
  that the resulting text is syntactically and
  morphologically correct.
• (Mellish and Dale, 1998)

7
Why Surface Realization?

• Relative agreement on
• input
• output
• goals
• More universally needed in MT than
• content planning
• text planning
• Somewhat easier to compare different techniques

8
Surface Realization Different Techniques

• Rule-based
• Templates
• Corpus-based
• Rules Corpus
• Rules Templates

9
Surface Realization Rule-Based

• Generates text using generation grammar rules
  (similar to rule-based understanding techniques)
• Most popular in the research community
• Long development time
• Often based on a specific linguistic theory
  (Systemic Functional Grammar seems most popular)
• Portable across domains (when grammar coverage is
  good)
• Systems FUF/SURGE, Penman, KPML

10
Surface Realization Rule-Based

• Input extensive semantic/syntactic features
  (e.g., from interlingua)
• Output high-quality sentences
• Knowledge Sources generation grammar,
  (domain-specific) lexicon
• Knowledge Acquisition hand-crafted
• Degradation from underspecified input default
  handling
• Degradation from lacking knowledge lower-quality
  output

11
Surface Realization Templates

• Generates text using canned expressions and
  hand-crafted templates
• Popular in commercial applications where similar
  documents are produced in large quantities (e.g.,
  customer service letter writing)
• Also popular in systems where generated output
  spans a narrow range (e.g., spoken dialog
  systems)
• Systems CLINT (business-letter writing)

12
Surface Realization Templates

• Input minimally specified
• Output limited set of sentences
• Knowledge Sources templates
• Knowledge Acquisition hand-crafted by looking at
  domain-specific corpora
• Degradation from underspecified input N/A
• Degradation from lacking knowledge no output

13
Surface Realization Corpus-based

• Developed for a task-oriented spoken dialog
  system
• Implemented in the CMU Communicator System
• Fast prototyping for different domains
• Natural output for spoken dialog
• A first attempt at a truly corpus-based
  stochastic generation
• Systems CMU, IBM

14
Surface Realization Corpus-Based

• Input dialog act
• Output medium-quality sentences
• Knowledge Sources language models
• Knowledge Acquisition domain-specific corpora
• Degradation from underspecified input N/A
• Degradation from lacking knowledge no output

15
Surface Realization Rules Corpus

• Nitrogen
• Developed to account for underspecified input
  which causes problems for rule-based techniques
• Developed for a machine translation project at
  ISI/USC
• Stochastic Generation at U. Edinburgh
• Accounts for sentence-level attributes (in
  addition to word-level attributes)
• An interesting technique to apply a certain
  authors style to another text (applying
  Shakespeares style, e.g., sentence length
  distribution and vocabulary diversity, to Mark
  Twain!)

16
Surface Realization Rules Corpus

• Input underspecified syntactic/semantic features
• Output high-quality sentences
• Knowledge Sources language models
• Knowledge Acquisition (domain-specific) corpora
• Degradation from underspecified input accounted
  for by language models
• Degradation from lacking knowledge lower-quality
  output

17
Surface Realization Rules Templates

• Enables more efficient development of NLG system
  for a concrete application
• Combines generation grammar rules, templates, and
  canned expressions
• Compensates for the shortcomings of each
  technique (i.e., utilizes the advantages of each
  technique)

18
Surface Realization Rules Templates

• Input underspecified syntactic/semantic features
• Output high-quality sentences
• Knowledge Sources generation grammar rules,
  lexicon, templates
• Knowledge Acquisition hand-crafted
• Degradation from underspecified input accounted
  for by templates and canned expressions
• Degradation from lacking knowledge lower-quality
  output

19
Future Directions

• Corpus-Based Techniques
• Langkilde, USC
• Multimodal (Multimedia) Generation
• McKeown, et al. Columbia University
• Concept-to-Speech Generation
• Hitzeman, et al. U. Edinburgh
• Hypertext (Web documents) Generation
• Dale, Macquarie University
• Reference Architecture for NLG
• RAGS Project (U. Edinburgh, U. Brighton)