Human Communication from the Perspective of Natural Language Processing

1
Human Communication from the Perspective of
Natural Language Processing

• Robert DaleRobert.Dale_at_mq.edu.au

2
The Aims of This Talk

• To provide an overview of how Natural Language
  Processing past and present views communication
• To identify existing connections with other
  disciplines in human communication science, and
  to seek new unexplored connections

3
Outline

• Some Definitions What Natural Language
  Processing is
• Different Perspectives on NLP
• The Conduit Metaphor in NLP
• NLP in Communication
• Narrowness

4
Some Definitions

• Natural Language Processing a blanket term used
  to cover a wide range of attempts to use
  computers to process natural language
• For many, includes speech as well as text (but
  typically not handwriting recognition)
• Includes both language analysis (input
  processing) and language generation (output
  processing)

5
Perspectives on NLP

• Computational Linguistics
• Natural Language Processing as a way of exploring
  formal models of language
• Computational Psycholinguistics
• Natural Language Processing as a way of exploring
  how the human language processing mechanism works
• Language Technology
• Natural Language Processing as the engineering of
  language sensitive applications

6
Key Problems in NLP

• Traditional Problems
• Word sense disambiguation, parsing, anaphora
  resolution, discourse structure,
• Current foci
• Named entity recognition, cross-document
  co-reference, semantic role labelling,

7
Language Technology Applications

• Spoken Language Dialog Systems
• Machine Translation
• Text Summarisation
• Question-answering Systems
• Grammar Checking
• Natural Language Generation
• Handwriting recognition, optical character
  recognition?
• Information retrieval?

8
Communication in Natural Language Processing The
Conduit Metaphor
9
Communication in Natural Language Processing The
Conduit Metaphor
Perceptions
Knowledge Base
User Model
Communicative Goal
Reasoning Agent
Text Plan
Disambiguation
Choice
Semantc Content
Syntactic Structure
Linearisation
10
Representations and Sources of Ideasin NLPs
Conduit Metaphor
11
Key Points

• Most NLP work is based on an implicit working out
  of the conduit metaphor
• Communication is seen as an intentional activity
  on the part of an agent who wants to communicate
  a message to an audience
• NLP is concerned with reasoning over
  representations
• A focus on computational modelling, and therefore
  formal representation this is open to the
  criticism of oversimplification

12
Modelling CommunicationSpoken Language Dialog
Systems
Reasoning Agent
Reasoning Agent
Communicative Goal
Communicative Goal
Text Plan
Text Plan
Semantic Content
Semantic Content
Syntactic Structure
Syntactic Structure
Linearisation
Linearisation
13
Modelling CommunicationNatural Language
Generation
Reasoning Agent
Communicative Goal
Text Plan
Semantic Content
Syntactic Structure
Linearisation
14
Modelling CommunicationMachine Translation
Reasoning Agent
Reasoning Agent
Communicative Goal
Communicative Goal
Text Plan
Text Plan
Semantic Content
Semantic Content
Syntactic Structure
Syntactic Structure
Linearisation
Linearisation
15
Modelling CommunicationText Summarisation
Reasoning Agent
Reasoning Agent
Communicative Goal
Communicative Goal
Text Plan
Text Plan
Semantic Content
Semantic Content
Summarisation
Syntactic Structure
Syntactic Structure
Linearisation
Linearisation
16
Modelling CommunicationText Summarisation
Reasoning Agent
Reasoning Agent
Communicative Goal
Communicative Goal
Text Plan
Text Plan
Semantic Content
Semantic Content
Syntactic Structure
Syntactic Structure
Summarisation
Linearisation
Linearisation
17
Modelling CommunicationGrammar Checking
Reasoning Agent
Communicative Goal
Text Plan
Semantic Content
Syntactic Structure
Linearisation
18
The Role of NLP in Communication

• Machine as conversational participant
• Extract meaning from communicative act, do
  something with it, respond via another
  communicative act
• Machine as speaker
• Communicate the content of an information source
• Machine as eavesdropper
• Do something useful with the results of someone
  elses communicative acts
• Machine as intermediary
• Convey the content of one agents communicative
  act to another

19
Narrowness

• 1960s-1980s Propositional Meaning
• A focus on information content, affective aspects
  of communication seen as secondary goal of both
  deep and broad coverage
• 1980s-1990s Skeletal Meaning
• Ignore most of a text the meaning consists of
  a few salient and important elements
• 1990s-2000s Textual Meaning
• No abstract representations, everything is
  captured in terms of statistics over text
  co-occurences

20
Textual Meaning

• Simplest case
• Summarisation by sentence extraction
• More sophisticated
• Statistical machine translation
• The t-shirt is white. ? Raha cumbra asto cocho.
• The t-shirt is black. ? Raha siembra asto cocho.

21
Textual Entailment

• Text
• Names in the News After Sweden swept three tennis
  matches to capture the Davis Cup, Yevgeny
  Kafelnikov restored a little Russian pride by
  defeating Stefan Edberg, 4-6, 6-4, 6-0, in
  Moscow.
• Hypothesis
• Sweden won the Davis Cup.

22
Two Alternate Views of the Way Forward

• Reasoning over logical representations is the
  right way to go we just havent worked out how
  to get there yet, and in the interim well see
  how far we can get with textual approximations
  to reasoning
• The whole idea of reasoning over logical
  representations is just a convenient abstraction
  for us to use in thinking about and discussing
  language processing, but it has nothing to do
  with real communication